CCAIR
CCAIR Student Guide · Part 6 of 8

Find data for cybersecurity projects

Every dataset has flaws. This catalog names them next to each entry, so you can choose data that fits your question and report its limits honestly.

Last verified: October 2026 · Reviewed by Dr. Benjamin M. Ampel, Director of CCAIR

Use a dataset responsibly

  • Read the caveats first. Every entry below names at least one known weakness. Report the dataset version and its weaknesses in your paper.
  • Cite the paper the creators ask for. Each entry links it.
  • Check access and license terms. Some datasets need a request or registration, and some limit commercial use.
  • Get approval for data about people. See the ethics section of Do research.

Public frameworks to pair with your data

  • MITRE ATT&CK. A knowledge base of attacker tactics and techniques, useful for labeling attacks[1].
  • CVE. The public list of identifiers for disclosed vulnerabilities[2].
  • National Vulnerability Database (NVD). NIST's database that adds severity scores and details to CVE records[3].

Dataset catalog

246 datasets in 13 categories. Search by name, topic or data type, or filter by category.

Showing 246 datasets.

  • AdvBench

    AI and LLM securityAdversarial suffix attacks

    AdvBench accompanies the Greedy Coordinate Gradient attack paper on aligned large language models (LLMs). It lists harmful behaviors and harmful strings used to test adversarial suffixes.

    Known weaknesses

    Many behaviors are near duplicates, so later papers often use a filtered subset. Prompts contain harmful requests by design and may produce dangerous model outputs. Store them securely and follow the license and responsible use terms.

    Access
    Open download
    License
    MIT
    Data type
    Harmful behaviors and target strings (CSV)
    Size
    520 behaviors and 574 strings
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for AdvBenchPaper to cite for AdvBench

  • AgentDojo

    AI and LLM securityAgent prompt injection

    AgentDojo is a dynamic environment for testing prompt injection against large language model (LLM) agents. It has 97 realistic user tasks and 629 security test cases.

    Known weaknesses

    Running the benchmark calls commercial model interfaces and can incur usage costs. Results depend on the agent scaffold and the defenses that are enabled.

    Access
    Open download
    License
    MIT
    Data type
    Python environment with tasks, tools, and attacks
    Size
    97 tasks and 629 security cases
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for AgentDojoPaper to cite for AgentDojo

  • AgentHarm

    AI and LLM securityHarmful agent behavior

    The UK AI Security Institute and Gray Swan released AgentHarm to measure harmful behavior by large language model (LLM) agents. It covers 110 malicious agent tasks across 11 harm categories including fraud and cybercrime.

    Known weaknesses

    The authors ask users not to train models on the benchmark. Prompts contain harmful requests by design and may produce dangerous model outputs. Store them securely and follow the license and responsible use terms.

    Access
    Open download
    License
    MIT with added usage clause
    Data type
    Agent tasks with synthetic tools and grading rubrics
    Size
    110 tasks, 440 with augmentations
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for AgentHarmPaper to cite for AgentHarm

  • AI Incident Database

    AI and LLM securityAI harm incidents

    The Responsible AI Collaborative catalogs real world incidents in which artificial intelligence (AI) systems caused or nearly caused harm. Weekly snapshots of the full database are available as downloadable archives.

    Known weaknesses

    Entries depend on media reports and on the work of volunteer editors. Coverage and classification can be uneven across regions and harm types.

    Access
    Open download
    License
    Not stated
    Data type
    Incident reports in JSON, CSV, and MongoDB dumps
    Size
    About 100 MB per recent snapshot
    Time span
    Weekly snapshots
    Link checked
    2026-09-27

    Dataset page for AI Incident DatabasePaper to cite for AI Incident Database

  • AI Vulnerability Database (AVID)

    AI and LLM securityAI vulnerability reports

    The AI Vulnerability Database (AVID) records failure modes and vulnerabilities of machine learning models. Its GitHub repository stores reports and vulnerability entries in structured files.

    Known weaknesses

    The repository shows limited recent activity compared with active incident trackers. Check entry dates before relying on the database for current coverage.

    Access
    Open download
    License
    MIT
    Data type
    Structured JSON reports and vulnerabilities
    Size
    Not stated
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for AI Vulnerability Database (AVID)

  • BIPIA

    AI and LLM securityIndirect prompt injection

    Microsoft built the Benchmark for Indirect Prompt Injection Attacks (BIPIA) on large language models (LLMs). Tasks cover email, web, and table question answering, summarization, and code.

    Known weaknesses

    The repository was archived in September 2026 and is now read only. Web and summarization data must be fetched from original sources.

    Access
    Open download
    License
    See repository LICENSE
    Data type
    Task contexts, attack instructions, and evaluation code
    Size
    Not stated
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for BIPIAPaper to cite for BIPIA

  • Cybench

    AI and LLM securityProfessional capture the flag evaluation

    Cybench evaluates language model agents on 40 professional capture the flag (CTF) tasks from four competitions. Many tasks include subtasks that break the challenge into steps for finer scoring.

    Known weaknesses

    Tasks run vulnerable services inside containers on the host machine. Use an isolated environment and review model outputs before execution.

    Access
    Open download
    License
    Apache 2.0
    Data type
    Task containers, subtasks, and agent harness
    Size
    40 tasks
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for CybenchPaper to cite for Cybench

  • CyberBench (JPMorgan Chase)

    AI and LLM securityCybersecurity natural language tasks

    CyberBench groups ten datasets to evaluate large language models (LLMs) on cybersecurity language tasks. Tasks include entity recognition, summarization, multiple choice, and text classification.

    Known weaknesses

    The repository was archived in May 2026 and is now read only. Source datasets keep their own licenses, which you must check separately.

    Access
    Open download
    License
    Apache 2.0
    Data type
    Scripts that assemble a unified CSV from source datasets
    Size
    Ten datasets
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for CyberBench (JPMorgan Chase)

  • deepset Prompt Injections

    AI and LLM securityPrompt injection classification

    deepset published a labeled set of benign and prompt injection inputs for training detection classifiers. It includes English and German examples split into training and test sets.

    Known weaknesses

    The set is small and many injections are simple and easy to detect. Test detectors on harder benchmarks before using them in deployed systems.

    Access
    Open download
    License
    Apache 2.0
    Data type
    Text with binary labels
    Size
    662 prompts
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for deepset Prompt Injections

  • Gandalf Ignore Instructions

    AI and LLM securityPrompt injection attempts

    Lakera released prompt injection attempts collected from its Gandalf password game. The prompts try to make a model ignore its instructions and reveal a secret.

    Known weaknesses

    All samples are attacks, so pair them with benign prompts for classifier training. Some prompts may include offensive language written by game players.

    Access
    Open download
    License
    MIT
    Data type
    Prompt text with similarity scores
    Size
    1,000 prompts
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for Gandalf Ignore Instructions

  • HarmBench

    AI and LLM securityJailbreak and red teaming benchmark

    HarmBench is a standardized framework for evaluating automated red teaming of large language models (LLMs). The authors compared 18 red teaming methods against 33 target models and defenses.

    Known weaknesses

    Prompts contain harmful requests by design and may produce dangerous model outputs. Store them securely and follow the license and responsible use terms.

    Access
    Open download
    License
    MIT
    Data type
    Harmful behavior prompts, classifier, and evaluation code
    Size
    510 behaviors (text and multimodal)
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for HarmBenchPaper to cite for HarmBench

  • In-the-Wild Jailbreak Prompts

    AI and LLM securityReal-world jailbreak prompts

    This dataset collects 15,140 prompts shared on Reddit, Discord, websites, and open datasets. Of these, 1,405 are jailbreak prompts aimed at large language models (LLMs).

    Known weaknesses

    Prompts contain harmful requests by design and may produce dangerous model outputs. Store them securely and follow the license and responsible use terms.

    Access
    Open download
    License
    MIT
    Data type
    Prompt text with platform and date metadata
    Size
    15,140 prompts plus 390 test questions
    Time span
    December 2022 to December 2023
    Link checked
    2026-09-27

    Dataset page for In-the-Wild Jailbreak PromptsPaper to cite for In-the-Wild Jailbreak Prompts

  • InjecAgent

    AI and LLM securityAgent indirect prompt injection

    InjecAgent benchmarks indirect prompt injection against tool integrated large language model (LLM) agents. It contains 1,054 test cases with 17 user tools and 62 attacker tools.

    Known weaknesses

    Tool responses are simulated rather than executed in real applications. Real deployments with live tools may behave differently from these results.

    Access
    Open download
    License
    MIT
    Data type
    JSON test cases and evaluation code
    Size
    1,054 test cases
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for InjecAgentPaper to cite for InjecAgent

  • InterCode-CTF

    AI and LLM securityInteractive capture the flag

    InterCode is a framework for interactive coding with execution feedback. Its capture the flag (CTF) split uses picoCTF problems solved inside a Bash environment.

    Known weaknesses

    Recent models solve most tasks, so the split now shows ceiling effects. Pair it with harder benchmarks such as Cybench or NYU CTF Bench.

    Access
    Open download
    License
    MIT
    Data type
    Tasks with Docker environments
    Size
    About 100 picoCTF tasks
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for InterCode-CTFPaper to cite for InterCode-CTF

  • JailbreakBench

    AI and LLM securityJailbreak benchmark

    JailbreakBench is an open robustness benchmark for jailbreaking large language models (LLMs). It pairs 100 harmful behaviors with 100 benign behaviors and a public leaderboard.

    Known weaknesses

    Prompts contain harmful requests by design and may produce dangerous model outputs. Store them securely and follow the license and responsible use terms.

    Access
    Open download
    License
    MIT
    Data type
    Behavior prompts, jailbreak artifacts, and evaluation library
    Size
    200 behaviors
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for JailbreakBenchPaper to cite for JailbreakBench

  • NYU CTF Bench

    AI and LLM securityCapture the flag evaluation

    NYU CTF Bench provides dockerized capture the flag (CTF) challenges from CSAW competitions for testing language model agents. Categories include web, pwn, forensics, reverse engineering, crypto, and misc.

    Known weaknesses

    Challenges contain exploitable services that could be misused if exposed. Run them only in isolated containers on a controlled network.

    Access
    Open download
    License
    GPL-2.0
    Data type
    Dockerized challenges with metadata
    Size
    200 test and 55 development challenges
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for NYU CTF BenchPaper to cite for NYU CTF Bench

  • RobustBench

    AI and LLM securityAdversarial robustness leaderboard

    RobustBench is a standardized benchmark and model zoo for adversarial robustness of image classifiers. It tracks certified and empirical robustness on CIFAR and ImageNet.

    Known weaknesses

    The benchmark covers image classifiers, not large language models or text systems. Pretrained model licenses vary by entry, so check each before reuse.

    Access
    Open download
    License
    See repository LICENSE
    Data type
    Pretrained models, leaderboard entries, and evaluation code
    Size
    Not stated
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for RobustBenchPaper to cite for RobustBench

  • SecBench

    AI and LLM securityCybersecurity knowledge and reasoning

    SecBench evaluates large language models (LLMs) on cybersecurity with questions in Chinese and English. It spans nine subdomains such as data, network, application, and cloud security.

    Known weaknesses

    Only a subset of the full question pool is public. The repository states no license, so confirm reuse terms with the authors.

    Access
    Open download
    License
    Not stated
    Data type
    Multiple choice and short answer questions (JSONL)
    Size
    3,000 public questions from a pool of 47,910
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for SecBenchPaper to cite for SecBench

  • SecQA

    AI and LLM securityComputer security questions

    SecQA is a small multiple choice benchmark on computer security concepts for large language models (LLMs). GPT-4 generated the questions using content from a computer security textbook.

    Known weaknesses

    The set is small and machine generated, which limits statistical power. Treat scores as a quick check rather than a full evaluation.

    Access
    Open download
    License
    CC BY-NC-SA 4.0
    Data type
    Multiple choice questions
    Size
    242 questions in two versions
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for SecQAPaper to cite for SecQA

  • Tensor Trust

    AI and LLM securityPrompt injection game data

    Tensor Trust is an online game where players attack and defend large language model (LLM) prompts. The dataset contains over 126,000 prompt injection attacks and 46,000 defenses.

    Known weaknesses

    The repository states no license, so confirm reuse terms with the authors. Player-written prompts may contain offensive content, so screen samples before sharing them.

    Access
    Open download
    License
    Not stated
    Data type
    Attack and defense prompts with benchmark splits
    Size
    126,000+ attacks and 46,000+ defenses
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for Tensor TrustPaper to cite for Tensor Trust

  • WildJailbreak

    AI and LLM securitySynthetic safety training data

    The Allen Institute for AI built WildJailbreak from jailbreak tactics mined from real chatbot interactions. It holds 262,000 vanilla and adversarial prompt and response pairs.

    Known weaknesses

    Access requires a Hugging Face account and acceptance of the dataset terms. Prompts contain harmful requests by design and may produce dangerous model outputs. Store them securely and follow the license and responsible use terms.

    Access
    Free registration
    License
    ODC-BY
    Data type
    Prompt and response pairs with harmful and benign labels
    Size
    About 262,000 pairs
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for WildJailbreakPaper to cite for WildJailbreak

  • WMDP

    AI and LLM securityHazardous knowledge evaluation

    The Weapons of Mass Destruction Proxy (WMDP) benchmark measures hazardous knowledge in language models. It has 3,668 multiple choice questions on biosecurity, cybersecurity, and chemical security.

    Known weaknesses

    Questions were filtered by experts to exclude sensitive and hazardous details. Use the set only for evaluation and unlearning research as the authors intend.

    Access
    Open download
    License
    MIT
    Data type
    Multiple choice questions
    Size
    3,668 questions
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for WMDPPaper to cite for WMDP

  • Awesome Cybersecurity Datasets (shramos)

    Dataset indexes and meta-repositoriesGitHub curated list

    This GitHub list gathers links to public cybersecurity datasets by topic. Topics include network traffic, malware, phishing, spam, fraud, and other areas.

    Known weaknesses

    The list was last updated in 2024 and some links may be broken. Verify each dataset at its original source before citing or using it.

    Access
    Open download
    License
    Not stated
    Data type
    Curated link list
    Size
    Not stated
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for Awesome Cybersecurity Datasets (shramos)

  • Awesome Machine Learning for Cyber Security (jivoi)

    Dataset indexes and meta-repositoriesGitHub curated list

    This GitHub list collects datasets, papers, and tools for machine learning in cybersecurity. Sections cover phishing, spam, malware, network intrusion, fraud, and related topics.

    Known weaknesses

    The list is not updated often and some entries are outdated. Many dataset links are older and may no longer be available.

    Access
    Open download
    License
    See repository LICENSE
    Data type
    Curated link list
    Size
    Not stated
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for Awesome Machine Learning for Cyber Security (jivoi)

  • Awesome-LM-SSP

    Dataset indexes and meta-repositoriesAI safety, security, and privacy reading list

    Awesome-LM-SSP is a curated list of papers and resources on safety, security, and privacy of large models. It also includes separate sections for benchmarks, toolkits, and public leaderboards.

    Known weaknesses

    The list focuses on research papers rather than on released datasets. Follow each listed paper to find its data release and license.

    Access
    Open download
    License
    Apache 2.0
    Data type
    Curated paper and resource list
    Size
    About 2,400 papers
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for Awesome-LM-SSP

  • Cambridge Cybercrime Centre Datasets

    Dataset indexes and meta-repositoriesCybercrime data index

    The Cambridge Cybercrime Centre shares cybercrime datasets with academic researchers under legal agreements. Holdings include underground forums, phishing, advance fee fraud, spam, and scam URLs.

    Known weaknesses

    Your institution must sign the data sharing agreement before any access is granted. Commercial use of the data is significantly restricted under the agreement terms.

    Access
    Academic request
    License
    Institutional data sharing agreement
    Data type
    Forum posts, emails, URLs, and chat messages
    Size
    Not stated
    Time span
    Varies by dataset, some from 2001
    Link checked
    2026-09-27

    Dataset page for Cambridge Cybercrime Centre Datasets

  • IMPACT Cyber Trust

    Dataset indexes and meta-repositoriesCybersecurity data repository

    IMPACT offers cybersecurity datasets and tools from multiple providers at no cost. Eligible researchers sign one memorandum of agreement with the coordinating center.

    Known weaknesses

    Eligibility is limited to researchers in the United States and listed partner countries. Some datasets are restricted and require extra review before release.

    Access
    Academic request
    License
    Memorandum of agreement per dataset class
    Data type
    Network, DNS, routing, and attack datasets
    Size
    Not stated
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for IMPACT Cyber Trust

  • Kaggle Cybersecurity Datasets

    Dataset indexes and meta-repositoriesCommunity dataset search

    Kaggle hosts many community uploaded cybersecurity datasets that are searchable by keyword and tag. Official competition data such as IEEE-CIS Fraud Detection also appears in the index.

    Known weaknesses

    Many uploads are mirrors without clear provenance or license information. Prefer datasets posted by the original authors or their organizations.

    Access
    Free registration
    License
    Varies by dataset
    Data type
    Search index of datasets
    Size
    Not stated
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for Kaggle Cybersecurity Datasets

  • SecRepo

    Dataset indexes and meta-repositoriesCurated security data index

    SecRepo is a curated list of security data samples maintained by Mike Sconzo. It covers network, malware, system, and threat intelligence data from many sources.

    Known weaknesses

    Many links point to third party sites that may change or disappear. Check each source for current availability, documentation, and license terms.

    Access
    Open download
    License
    CC BY 4.0 for hosted data, exceptions noted
    Data type
    Links and hosted samples
    Size
    Not stated
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for SecRepo

  • Stratosphere Laboratory Datasets

    Dataset indexes and meta-repositoriesMalware and network traffic index

    The Stratosphere Laboratory at Czech Technical University hosts malware, normal, and mixed network captures. Collections include CTU-13, IoT-23, Hornet, and the Android Mischief dataset.

    Known weaknesses

    Some captures contain live malware traffic and embedded malicious payloads. Analyze packet captures only on isolated systems without network access.

    Access
    Open download
    License
    CC BY (CTU-13)
    Data type
    Packet captures and bidirectional flow files
    Size
    Not stated
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for Stratosphere Laboratory Datasets

  • UNB Canadian Institute for Cybersecurity (CIC) Datasets

    Dataset indexes and meta-repositoriesLab dataset index

    The University of New Brunswick Canadian Institute for Cybersecurity (CIC) lists its public research datasets. Topics include intrusion detection, malware, Internet of Things, DNS, URLs, and language models.

    Known weaknesses

    Each download passes through a short form that collects basic user details. Several older datasets have known labeling issues reported in the literature.

    Access
    Free registration
    License
    Redistribution allowed with citation of dataset and paper
    Data type
    Index of datasets with download forms
    Size
    Not stated
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for UNB Canadian Institute for Cybersecurity (CIC) Datasets

  • Zenodo Communities

    Dataset indexes and meta-repositoriesResearch data repository

    Zenodo is a CERN research repository that hosts many security datasets with persistent identifiers. Communities group related records so users can browse collections by topic.

    Known weaknesses

    Quality and documentation vary widely by uploader and by community. Check the linked paper and license for each record before use.

    Access
    Open download
    License
    Varies by record
    Data type
    Search index of records
    Size
    Not stated
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for Zenodo Communities

  • ASVspoof 2019

    Deepfakes and voice spoofingVoice spoofing detection

    The Automatic Speaker Verification Spoofing (ASVspoof) 2019 database covers synthesized, converted, and replayed speech. It includes logical access and physical access scenarios with separate protocols.

    Known weaknesses

    The speech is studio quality and comes from about 100 speakers. Models trained on it may not generalize to real world recordings.

    Access
    Open download
    License
    ODC-BY
    Data type
    Audio (FLAC) with protocol files
    Size
    Not stated
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for ASVspoof 2019Paper to cite for ASVspoof 2019

  • ASVspoof 2021

    Deepfakes and voice spoofingVoice spoofing and deepfake detection

    The Automatic Speaker Verification Spoofing (ASVspoof) 2021 challenge added telephony channels and a speech deepfake task. Evaluation data for logical access, physical access, and deepfake tasks is hosted on Zenodo.

    Known weaknesses

    Training data comes from the ASVspoof 2019 release rather than a new set. Download both releases to run complete training and evaluation experiments.

    Access
    Open download
    License
    ODC-BY
    Data type
    Audio (FLAC) with keys and metadata
    Size
    Logical access evaluation set about 7.8 GB
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for ASVspoof 2021Paper to cite for ASVspoof 2021

  • ASVspoof 5

    Deepfakes and voice spoofingVoice spoofing and adversarial attacks

    The Automatic Speaker Verification Spoofing (ASVspoof) 5 database uses crowdsourced speech from about 2,000 speakers. It includes spoofing attacks from 32 algorithms plus seven adversarial attacks.

    Known weaknesses

    The archives are large and need substantial storage and bandwidth. Use the official protocols to keep speaker partitions disjoint across splits.

    Access
    Open download
    License
    ODC-BY
    Data type
    Audio (FLAC) with protocol files
    Size
    About 150 GB across many archives
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for ASVspoof 5Paper to cite for ASVspoof 5

  • Deepfake Detection Challenge (DFDC)

    Deepfakes and voice spoofingVideo deepfake detection

    Meta released the Deepfake Detection Challenge (DFDC) dataset of face swap videos made with paid actors. The full set has over 100,000 clips from 3,426 actors.

    Known weaknesses

    Download requires an AWS account and acceptance of the data terms. The full set is very large and needs substantial storage and bandwidth.

    Access
    Free registration
    License
    DFDC terms of use (research)
    Data type
    Video clips with labels
    Size
    About 124,000 videos
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for Deepfake Detection Challenge (DFDC)Paper to cite for Deepfake Detection Challenge (DFDC)

  • FaceForensics++

    Deepfakes and voice spoofingFacial manipulation detection

    FaceForensics++ contains 1,000 original videos and their manipulated versions from four face manipulation methods. It also includes the Google DeepFakeDetection subset with paid actors.

    Known weaknesses

    Access requires a Google form and acceptance of the terms of use. Source videos were downloaded from YouTube, so respect the platform terms of use.

    Access
    Academic request
    License
    FaceForensics terms of use
    Data type
    Video sequences, masks, and model files
    Size
    1,000 source videos with four manipulations
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for FaceForensics++Paper to cite for FaceForensics++

  • FakeAVCeleb

    Deepfakes and voice spoofingAudio-video deepfake detection

    FakeAVCeleb pairs deepfake videos with cloned audio for multimodal detection research. It contains 500 real videos and 19,500 fake videos with matching audio.

    Known weaknesses

    Access requires a Google form and approval by the authors. Subjects are celebrities who did not consent to the manipulation of their likeness.

    Access
    Academic request
    License
    Research use under request terms
    Data type
    Video with synchronized audio
    Size
    20,000 videos
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for FakeAVCelebPaper to cite for FakeAVCeleb

  • In-the-Wild Audio Deepfake

    Deepfakes and voice spoofingAudio deepfake detection

    This dataset holds real and deepfake audio of 58 politicians and public figures collected online. It tests whether audio deepfake detectors generalize to realistic recordings.

    Known weaknesses

    Clips depict real public figures, so handle them with care in publications. Do not use these recordings to train voice cloning or impersonation systems.

    Access
    Open download
    License
    Apache 2.0 per project page, CC BY-SA 4.0 per Hugging Face card
    Data type
    Audio clips with labels
    Size
    20.8 hours real and 17.2 hours fake
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for In-the-Wild Audio DeepfakePaper to cite for In-the-Wild Audio Deepfake

  • WaveFake

    Deepfakes and voice spoofingAudio deepfake detection

    WaveFake contains about 175 hours of synthesized speech from several neural vocoder architectures. It supports research on detecting audio produced by neural speech synthesis systems.

    Known weaknesses

    Fake audio comes from only two source speakers in English and Japanese. Detectors may learn speaker traits instead of the artifacts of synthesis.

    Access
    Open download
    License
    CC BY-SA 4.0 (data), MIT (code)
    Data type
    Audio (WAV) generated from LJSpeech and JSUT
    Size
    About 29 GB
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for WaveFakePaper to cite for WaveFake

  • Bank Account Fraud (BAF) Dataset Suite

    Fraud and financial crimeAccount opening fraud

    Feedzai released six synthetic bank account opening datasets for the Neural Information Processing Systems (NeurIPS) 2022 conference. The suite supports research on fairness, class imbalance, and changing data distributions.

    Known weaknesses

    The data is generated from a real dataset with privacy controls. Protected attributes such as age are included to support bias studies.

    Access
    Free registration
    License
    CC BY-NC-SA 4.0
    Data type
    Tabular synthetic applications (CSV)
    Size
    Six datasets of one million rows each, about 1.4 GB
    Time span
    Eight simulated months
    Link checked
    2026-09-27

    Dataset page for Bank Account Fraud (BAF) Dataset SuitePaper to cite for Bank Account Fraud (BAF) Dataset Suite

  • Credit Card Fraud Detection (ULB)

    Fraud and financial crimeCard transaction fraud

    The Machine Learning Group at Universite Libre de Bruxelles (ULB) released European card transactions from two days. Only 492 of the 284,807 transactions are fraudulent, giving a highly imbalanced task.

    Known weaknesses

    Most features are anonymized components, so feature engineering is limited. Severe class imbalance requires precision and recall metrics instead of plain accuracy.

    Access
    Free registration
    License
    Open Database License (ODbL) with Database Contents License
    Data type
    Tabular features from principal component analysis, plus time and amount
    Size
    284,807 transactions, about 150 MB
    Time span
    Two days in September 2013
    Link checked
    2026-09-27

    Dataset page for Credit Card Fraud Detection (ULB)Paper to cite for Credit Card Fraud Detection (ULB)

  • DGraphFin

    Fraud and financial crimeFinancial fraud graph

    DGraphFin is a dynamic social graph from the Finvolution Group for graph anomaly detection. It has about 3 million nodes, 4 million dynamic edges, and 1 million labeled nodes.

    Known weaknesses

    The download requires a free account on the DGraph website. Node features are anonymized, which limits interpretation of model results.

    Access
    Free registration
    License
    Not stated
    Data type
    Dynamic graph with node features and labels
    Size
    About 3 million nodes
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for DGraphFinPaper to cite for DGraphFin

  • Elliptic Data Set

    Fraud and financial crimeCryptocurrency money laundering

    Elliptic released a Bitcoin transaction graph with 203,769 nodes and 234,355 edges. About 2 percent of transactions are labeled illicit and 21 percent licit.

    Known weaknesses

    Most nodes are unlabeled, so semi-supervised methods are common with this graph. The license forbids commercial use and redistribution of derivative versions.

    Access
    Free registration
    License
    CC BY-NC-ND 4.0
    Data type
    Graph edge list and node features with labels (CSV)
    Size
    203,769 transactions, about 700 MB
    Time span
    49 time steps
    Link checked
    2026-09-27

    Dataset page for Elliptic Data SetPaper to cite for Elliptic Data Set

  • Elliptic++

    Fraud and financial crimeCryptocurrency fraud and illicit actors

    Elliptic++ extends the Elliptic dataset with 822,942 Bitcoin wallet addresses and their interactions. It supports detection of both illicit transactions and illicit actors.

    Known weaknesses

    The files are hosted on Google Drive through a repository link. No license is stated in the repository, so confirm terms with the authors.

    Access
    Open download
    License
    Not stated
    Data type
    Transaction and wallet graphs with features (CSV)
    Size
    203,769 transactions and 822,942 addresses
    Time span
    49 time steps
    Link checked
    2026-09-27

    Dataset page for Elliptic++Paper to cite for Elliptic++

  • Elliptic2

    Fraud and financial crimeCryptocurrency money laundering subgraphs

    Elliptic2 provides 122,000 labeled subgraphs of Bitcoin clusters inside a background graph. The background graph holds 49 million node clusters and 196 million edges.

    Known weaknesses

    The dataset is very large and needs substantial storage and memory. The license does not allow commercial use or redistribution of derivatives.

    Access
    Free registration
    License
    CC BY-NC-ND 4.0
    Data type
    Graph data with labeled subgraphs
    Size
    About 88 GB
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for Elliptic2Paper to cite for Elliptic2

  • Fraud Dataset Benchmark (FDB)

    Fraud and financial crimeFraud benchmark suite

    Amazon Science compiled the Fraud Dataset Benchmark (FDB) from nine public fraud related datasets. Its Python loaders give standard train and test splits and metrics.

    Known weaknesses

    Loaders pull data from Kaggle, so you need a Kaggle account and competition access. The repository has not been updated since 2023, so some loaders may break.

    Access
    Open download
    License
    MIT (code), source datasets keep their own licenses
    Data type
    Python loaders over tabular datasets
    Size
    Nine datasets
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for Fraud Dataset Benchmark (FDB)Paper to cite for Fraud Dataset Benchmark (FDB)

  • IBM Transactions for Anti-Money Laundering (AML)

    Fraud and financial crimeMoney laundering (synthetic)

    IBM researchers generated synthetic financial transactions with labeled laundering patterns for anti-money laundering (AML) research. Several dataset sizes and illicit ratios are available for different experiments.

    Known weaknesses

    The data is synthetic and the laundering typologies are predefined. Validate any findings on real transaction data where access is possible.

    Access
    Free registration
    License
    CDLA-Sharing 1.0
    Data type
    Tabular transactions and pattern files (CSV)
    Size
    About 42 GB across variants
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for IBM Transactions for Anti-Money Laundering (AML)Paper to cite for IBM Transactions for Anti-Money Laundering (AML)

  • IEEE-CIS Fraud Detection

    Fraud and financial crimeCard-not-present transaction fraud

    Vesta Corporation and the IEEE Computational Intelligence Society (IEEE-CIS) released this Kaggle competition dataset. It contains e-commerce transactions with identity features and fraud labels.

    Known weaknesses

    You must join the competition on Kaggle before the data can be downloaded. Many feature names are masked, which limits interpretation of model behavior.

    Access
    Free registration
    License
    Kaggle competition rules
    Data type
    Tabular transaction and identity tables (CSV)
    Size
    About 590,000 labeled training transactions
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for IEEE-CIS Fraud Detection

  • PaySim

    Fraud and financial crimeMobile money fraud (synthetic)

    PaySim is a synthetic mobile money dataset generated by an agent-based simulator. The simulator was calibrated on real transaction logs from an African mobile money service.

    Known weaknesses

    The data is simulated and fraud patterns follow simulator rules. Models may learn simulator artifacts that do not appear in real transactions.

    Access
    Free registration
    License
    CC BY-SA 4.0
    Data type
    Tabular synthetic transactions (CSV)
    Size
    About 6.3 million rows, about 490 MB
    Time span
    30 simulated days
    Link checked
    2026-09-27

    Dataset page for PaySim

  • AIT Log Data Set V2.0

    Host, endpoint, and security logsMulti-source host and security logs

    The Austrian Institute of Technology (AIT) Log Data Set V2.0 covers eight simulated enterprise testbeds. Each testbed logs several days of benign activity followed by multi-step attacks.

    Known weaknesses

    Benign activity is simulated by scripts and may lack real user variety. The noncommercial license bars use of the data in commercial products.

    Access
    Open download
    License
    CC BY-NC-SA 4.0
    Data type
    Apache, auth, DNS, VPN, audit, Suricata, and system logs, PCAP files
    Size
    130.6 GB compressed
    Time span
    4 to 6 days per testbed
    Link checked
    2026-09-27

    Dataset page for AIT Log Data Set V2.0Paper to cite for AIT Log Data Set V2.0

  • ATLAS

    Host, endpoint, and security logsAPT attack investigation logs

    ATLAS provides audit logs from multi-stage advanced persistent threat (APT) attacks on Windows hosts. Logs include Windows security events, Domain Name System (DNS) records, and browser activity.

    Known weaknesses

    Attacks were staged in a small lab with limited benign noise. The paper appeared at USENIX Security 2021 without a DOI.

    Access
    Open download
    License
    Apache-2.0
    Data type
    Windows security, DNS, and Firefox logs
    Size
    Not stated
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for ATLAS

  • ATLASv2

    Host, endpoint, and security logsAPT attack investigation logs

    ATLASv2 extends ATLAS with Sysmon and VMware Carbon Black Cloud logs. Researchers worked on the hosts for four days before running ten attack scenarios.

    Known weaknesses

    The hosts ran Windows 7, which is out of support. Many attack steps overlap across the ten scenarios, which reduces scenario diversity.

    Access
    Open download
    License
    Not stated
    Data type
    Sysmon, Carbon Black, Windows security, DNS, and Firefox logs
    Size
    Not stated
    Time span
    Five days per engagement
    Link checked
    2026-09-27

    Dataset page for ATLASv2Paper to cite for ATLASv2

  • Atomic Red Team

    Host, endpoint, and security logsAttack-simulation generator

    Atomic Red Team is a library of small attack tests mapped to MITRE ATT&CK. Teams run the tests to produce their own host telemetry.

    Known weaknesses

    This is a data generator and not a dataset of recorded logs. Tests change real systems and should run only in isolated lab machines.

    Access
    Open download
    License
    MIT
    Data type
    Attack test definitions (YAML) and scripts
    Size
    About 1,878 atomic tests
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for Atomic Red Team

  • BETH Dataset

    Host, endpoint, and security logsKernel process and DNS logs

    The Berkeley Packet Filter (BPF) extended tracking honeypot (BETH) dataset records kernel process calls from cloud honeypots. It also includes Domain Name System (DNS) logs, and each host shows at most one attack.

    Known weaknesses

    Download requires a free Kaggle account and acceptance of the dataset terms. Honeypot activity differs from the behavior of normal enterprise hosts.

    Access
    Free registration
    License
    Not stated
    Data type
    Kernel process logs and DNS logs (CSV)
    Size
    More than 8 million events from 23 hosts
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for BETH Dataset

  • CERT Insider Threat Test Dataset

    Host, endpoint, and security logsInsider threat activity logs

    The Carnegie Mellon University (CMU) CERT Division built synthetic logs of employee activity with injected insider threat scenarios. Files cover logons, devices, web, email, files, and personality scores.

    Known weaknesses

    All data is synthetic, and malicious insiders are very rare in each release. Release r6.2 has five malicious insiders among about 4,000 simulated users.

    Access
    Open download
    License
    Not stated
    Data type
    Synthetic CSV activity logs
    Size
    Multiple releases from r1 to r6.2
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for CERT Insider Threat Test DatasetPaper to cite for CERT Insider Threat Test Dataset

  • DARPA Transparent Computing (Engagements 3 and 5)

    Host, endpoint, and security logsAPT provenance and audit data

    The Defense Advanced Research Projects Agency (DARPA) Transparent Computing program recorded system provenance during red team exercises. Advanced persistent threat (APT) attacks ran against continuous benign activity.

    Known weaknesses

    Data sits on Google Drive and needs large storage and custom parsers. Ground truth is in prose reports rather than labeled events.

    Access
    Open download
    License
    Public domain
    Data type
    Provenance records in Common Data Model (CDM) Avro format
    Size
    Not stated
    Time span
    2018 (Engagement 3) and May 2019 (Engagement 5)
    Link checked
    2026-09-27

    Dataset page for DARPA Transparent Computing (Engagements 3 and 5)

  • EVTX-ATTACK-SAMPLES

    Host, endpoint, and security logsWindows event log samples

    This collection holds Windows event log (EVTX) files that capture specific attack techniques. Folders follow MITRE ATT&CK tactics to support detection testing and threat hunting.

    Known weaknesses

    Samples are short and isolated, so they lack benign baseline activity. The repository has not received new samples or updates since 2023.

    Access
    Open download
    License
    GPL-3.0
    Data type
    Windows EVTX files
    Size
    About 200 samples
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for EVTX-ATTACK-SAMPLES

  • HDFS Console Logs (SOSP 2009)

    Host, endpoint, and security logsSystem log anomaly detection

    This dataset holds Hadoop Distributed File System (HDFS) logs from a benchmark workload with block-level anomaly labels. It is a standard benchmark for log-based anomaly detection research.

    Known weaknesses

    Anomalies reflect system faults and performance problems rather than security attacks. Recent studies show simple methods reach near-perfect scores on it.

    Access
    Open download
    License
    BSD
    Data type
    Raw logs and labeled message count vectors
    Size
    About 11.2 million log lines
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for HDFS Console Logs (SOSP 2009)Paper to cite for HDFS Console Logs (SOSP 2009)

  • KYPO Cybersecurity Training Shell Commands

    Host, endpoint, and security logsCapture-the-flag and training telemetry

    This Masaryk University dataset records shell commands typed by trainees in KYPO cyber range exercises. Each training folder includes command logs, metadata, and the sandbox definition used.

    Known weaknesses

    The dataset is small and reflects student learners rather than attackers. It captures typed commands only and does not include full host telemetry.

    Access
    Open download
    License
    CC BY 4.0
    Data type
    Shell command logs (JSON)
    Size
    255 KB, seven trainings
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for KYPO Cybersecurity Training Shell CommandsPaper to cite for KYPO Cybersecurity Training Shell Commands

  • LANL User-Computer Authentication Associations in Time

    Host, endpoint, and security logsAuthentication logs

    This LANL dataset records successful user-to-computer authentication events across an enterprise network. It supports graph and anomaly detection research on enterprise credential use.

    Known weaknesses

    The data has no attack labels, so it suits unsupervised research. Each record has only time, user, and computer fields and no other context.

    Access
    Free registration
    License
    CC0-1.0
    Data type
    Authentication events (CSV)
    Size
    708,304,516 events, 2.3 GB compressed
    Time span
    9 months
    Link checked
    2026-09-27

    Dataset page for LANL User-Computer Authentication Associations in TimePaper to cite for LANL User-Computer Authentication Associations in Time

  • LID-DS

    Host, endpoint, and security logsSystem call intrusion detection

    The Leipzig Intrusion Detection Data Set (LID-DS) records system calls from services under normal use and attack. Attack scenarios exploit real published vulnerabilities in containerized server applications and services.

    Known weaknesses

    Downloads are hosted on Proton Drive rather than the repository. Scenarios run in containers and may not match full host behavior.

    Access
    Open download
    License
    GPL-3.0
    Data type
    System call traces with arguments and network captures
    Size
    Not stated
    Time span
    2019 and 2021 releases
    Link checked
    2026-09-27

    Dataset page for LID-DS

  • Loghub

    Host, endpoint, and security logsSystem log anomaly detection

    Loghub is a collection of system logs from distributed systems, supercomputers, operating systems, and applications. Several logs carry anomaly labels that support log analytics and detection research.

    Known weaknesses

    Only some datasets in the collection have anomaly labels for evaluation. Most logs describe system failures and faults rather than security attacks.

    Access
    Open download
    License
    Free for research with citation
    Data type
    Raw system logs (text)
    Size
    16 datasets, more than 150 GiB
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for LoghubPaper to cite for Loghub

  • Loghub-2.0

    Host, endpoint, and security logsLog parsing benchmark

    Loghub-2.0 provides large annotated log datasets for evaluating log parsing methods. It covers 14 systems with template labels for each line.

    Known weaknesses

    Annotations target log template extraction and not anomaly or attack detection.

    Access
    Open download
    License
    Free for research with citation
    Data type
    Annotated system logs
    Size
    About 3.6 million annotated lines per dataset on average
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for Loghub-2.0Paper to cite for Loghub-2.0

  • Microsoft GUIDE

    Host, endpoint, and security logsSIEM incident triage

    GUIDE is a Microsoft dataset of real security incidents with analyst triage labels. It spans alerts and evidence from more than 6,100 organizations.

    Known weaknesses

    Identifiers are scrubbed, so raw log content is not available. Download requires a free Kaggle account and acceptance of the dataset terms.

    Access
    Free registration
    License
    CDLA-Permissive-2.0
    Data type
    Tabular incidents, alerts, and evidence (CSV)
    Size
    About 13 million evidence rows, 1.6 million alerts, 1 million incidents
    Time span
    Two weeks
    Link checked
    2026-09-27

    Dataset page for Microsoft GUIDEPaper to cite for Microsoft GUIDE

  • OTRF Security-Datasets (formerly Mordor)

    Host, endpoint, and security logsAttack-simulation telemetry

    The Open Threat Research Forge (OTRF) Security-Datasets project shares logs from simulated adversary techniques. Datasets map to MITRE ATT&CK and include Windows events, Sysmon, and network captures.

    Known weaknesses

    Most datasets capture short attack runs with little benign background activity. Repository activity has slowed, with no new commits since early 2024.

    Access
    Open download
    License
    MIT
    Data type
    Windows event logs, Sysmon logs, PCAP files (JSON)
    Size
    Not stated
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for OTRF Security-Datasets (formerly Mordor)

  • Splunk Attack Data

    Host, endpoint, and security logsAttack-simulation telemetry

    Splunk Attack Data is a curated set of logs from simulated attacks mapped to MITRE ATT&CK. It supports detection rule development and testing for Splunk security content.

    Known weaknesses

    Files are stored with Git Large File Storage (LFS), so plain clones miss data. Each file covers a narrow technique with little benign context.

    Access
    Open download
    License
    Apache-2.0
    Data type
    Sysmon, Windows Security, EDR, and AWS logs
    Size
    About 9 GB
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for Splunk Attack Data

  • Splunk Boss of the SOC v1 (BOTSv1)

    Host, endpoint, and security logsSIEM and capture-the-flag telemetry

    Boss of the Security Operations Center (BOTS) v1 is a Splunk capture-the-flag (CTF) dataset. It combines Windows, Sysmon, firewall, intrusion detection, and network stream data from staged incidents.

    Known weaknesses

    Full use requires a Splunk installation and the matching technology add-ons. Challenge questions and answers sit in a separate Splunk scoreboard repository. The data may contain profanity or offensive language kept from the scenarios.

    Access
    Open download
    License
    CC0-1.0
    Data type
    Pre-indexed Splunk data, JSON, CSV
    Size
    6.1 GB indexed, 11.3 GB JSON
    Time span
    2016
    Link checked
    2026-09-27

    Dataset page for Splunk Boss of the SOC v1 (BOTSv1)

  • Splunk Boss of the SOC v2 (BOTSv2)

    Host, endpoint, and security logsSIEM and capture-the-flag telemetry

    BOTSv2 is a Splunk capture-the-flag (CTF) dataset with more than 90 source types. It covers Windows, Linux, Active Directory, firewalls, endpoint protection, and network traffic.

    Known weaknesses

    The data is only available in pre-indexed Splunk format for this release. The data may contain profanity or offensive language kept from the scenarios.

    Access
    Open download
    License
    CC0-1.0
    Data type
    Pre-indexed Splunk data
    Size
    16.4 GB full or 3.2 GB attack-only
    Time span
    2017
    Link checked
    2026-09-27

    Dataset page for Splunk Boss of the SOC v2 (BOTSv2)

  • Splunk Boss of the SOC v3 (BOTSv3)

    Host, endpoint, and security logsSIEM and capture-the-flag telemetry

    BOTSv3 is a Splunk capture-the-flag (CTF) dataset with more than 100 source types. It adds cloud logs from Amazon Web Services (AWS), Azure, and Office 365.

    Known weaknesses

    It requires Splunk and more than 20 add-ons to parse all source types. Challenge questions are distributed separately from the dataset through the Splunk scoreboard.

    Access
    Open download
    License
    CC0-1.0
    Data type
    Pre-indexed Splunk data
    Size
    320.1 MB
    Time span
    2018
    Link checked
    2026-09-27

    Dataset page for Splunk Boss of the SOC v3 (BOTSv3)

  • StreamSpot

    Host, endpoint, and security logsProvenance graphs

    StreamSpot provides provenance graphs from benign browsing and application scenarios and one drive-by download attack. It supports research on streaming graph anomaly detection for host intrusion.

    Known weaknesses

    Detection performance on this dataset is largely saturated in recent studies. The data dates from 2016 and has only one attack scenario.

    Access
    Open download
    License
    Apache-2.0
    Data type
    Provenance graph edge lists
    Size
    600 graphs
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for StreamSpotPaper to cite for StreamSpot

  • Unraveled

    Host, endpoint, and security logsAPT host and network data

    Unraveled is a semi-synthetic dataset that follows an advanced persistent threat (APT) campaign from both views. It combines host logs and network flows from a cloud-based emulated organization.

    Known weaknesses

    User behavior is scripted, so benign activity may be more regular than real traffic.

    Access
    Open download
    License
    GPL-3.0
    Data type
    Host logs and network flows
    Size
    Not stated
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for UnraveledPaper to cite for Unraveled

  • USENIX CFDR HPC4 Logs (BGL and Thunderbird)

    Host, endpoint, and security logsSystem log anomaly detection

    The Computer Failure Data Repository (CFDR) hosts supercomputer logs from Blue Gene/L (BGL), Thunderbird, and three other systems. Messages carry alert category tags that researchers use as anomaly labels.

    Known weaknesses

    Alerts describe hardware and software faults and do not represent attacks. The logs are nearly 20 years old and come from retired systems.

    Access
    Open download
    License
    Not stated
    Data type
    Raw system logs with alert tags
    Size
    BGL 60 MB and Thunderbird 1.9 GB compressed
    Time span
    2004 to 2006
    Link checked
    2026-09-27

    Dataset page for USENIX CFDR HPC4 Logs (BGL and Thunderbird)Paper to cite for USENIX CFDR HPC4 Logs (BGL and Thunderbird)

  • 4SICS ICS Lab PCAPs

    IoT, ICS, and OT securityICS and SCADA

    Netresec captured network traffic in the ICS lab at the 4SICS conference in Stockholm. The captures include Modbus, DNP3, and other industrial protocols from real control devices.

    Known weaknesses

    The captures are unlabeled and come from a conference lab. They suit protocol analysis and teaching more than benchmark evaluation.

    Access
    Open download
    License
    Credit to CS3Sthlm requested
    Data type
    PCAP
    Size
    About 359 MB across three files
    Time span
    October 20-22, 2015
    Link checked
    2026-09-27

    Dataset page for 4SICS ICS Lab PCAPs

  • BATADAL

    IoT, ICS, and OT securityICS and SCADA

    The Battle of the Attack Detection Algorithms (BATADAL) data simulates SCADA readings from the C-Town water distribution network. It provides attack-free training data, partially labeled attack data, and a test set.

    Known weaknesses

    All data is simulated with hydraulic modeling rather than measured from a real utility. Some attacks in the second training set are only approximately labeled.

    Access
    Open download
    License
    CC BY 4.0
    Data type
    Time-series CSV
    Size
    Not stated
    Time span
    2016-2017 releases
    Link checked
    2026-09-27

    Dataset page for BATADALPaper to cite for BATADAL

  • BoT-IoT

    IoT, ICS, and OT securityIoT botnet traffic

    BoT-IoT contains normal and botnet traffic from a simulated IoT network in the UNSW Canberra Cyber Range. The attacks include DDoS, DoS, scanning, keylogging, and data exfiltration.

    Known weaknesses

    Attack records vastly outnumber normal records, which inverts typical class imbalance. Most IoT devices in the testbed were simulated rather than physical hardware.

    Access
    Open download
    License
    Free for academic research, commercial use requires author agreement
    Data type
    PCAP, Argus, CSV
    Size
    69.3 GB PCAP, 72M+ records, 16.7 GB CSV
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for BoT-IoTPaper to cite for BoT-IoT

  • Car-Hacking Dataset (HCRL)

    IoT, ICS, and OT securityVehicle networks

    The Hacking and Countermeasure Research Lab (HCRL) logged CAN traffic through the OBD-II port of a real vehicle. The data includes DoS, fuzzy, gear spoofing, and RPM spoofing injection attacks.

    Known weaknesses

    Data comes from a single vehicle, and injected messages arrive at fixed high rates such as every millisecond. The lab notes errors in Table 2 of the original PST paper.

    Access
    Open download
    License
    Academic use with citation
    Data type
    CAN message CSV
    Size
    Not stated
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for Car-Hacking Dataset (HCRL)Paper to cite for Car-Hacking Dataset (HCRL)

  • CIC IoT-DIAD 2024

    IoT, ICS, and OT securityIoT intrusion detection

    CIC IoT-DIAD 2024 supports both IoT device identification and anomaly detection from the same captures. It provides flow-based and packet-based features extracted from PCAP files.

    Known weaknesses

    Traffic comes from a lab testbed and may not match home or industrial deployments. Features come in two extraction formats, so results depend on the chosen format.

    Access
    Free registration
    License
    Not stated
    Data type
    PCAP, feature CSV
    Size
    Not stated
    Time span
    2024
    Link checked
    2026-09-27

    Dataset page for CIC IoT-DIAD 2024Paper to cite for CIC IoT-DIAD 2024

  • CIC Modbus Dataset 2023

    IoT, ICS, and OT securityICS and SCADA

    The CIC Modbus dataset contains network captures and logs from a simulated electrical substation using the Modbus protocol. Attacks come from external devices, a compromised intelligent electronic device (IED), and a compromised HMI.

    Known weaknesses

    The substation network is simulated rather than a live utility system. PCAP timestamps use Atlantic Daylight Time and must be aligned with log files.

    Access
    Free registration
    License
    Redistribution allowed with citation
    Data type
    PCAP, logs
    Size
    Not stated
    Time span
    2023
    Link checked
    2026-09-27

    Dataset page for CIC Modbus Dataset 2023Paper to cite for CIC Modbus Dataset 2023

  • CICIoMT2024

    IoT, ICS, and OT securityIoT intrusion detection

    CICIoMT2024 contains traffic from 18 attacks against a testbed of 40 Internet of Medical Things (IoMT) devices. It covers Wi-Fi, MQTT, and Bluetooth protocols for healthcare device security research.

    Known weaknesses

    Fifteen of the 40 devices are simulated rather than physical medical hardware. Attacks fall into five classes that differ widely in record counts.

    Access
    Free registration
    License
    Not stated
    Data type
    PCAP, feature CSV
    Size
    Not stated
    Time span
    2024
    Link checked
    2026-09-27

    Dataset page for CICIoMT2024Paper to cite for CICIoMT2024

  • CICIoT2023

    IoT, ICS, and OT securityIoT intrusion detection

    CICIoT2023 contains traffic from 33 attacks executed in an Internet of Things (IoT) topology of 105 devices. The seven attack classes are DDoS, DoS, reconnaissance, web-based, brute force, spoofing, and Mirai.

    Known weaknesses

    DDoS and DoS records dominate the class distribution of the released data. Attacks were run in a lab, so benign traffic is limited in volume and variety.

    Access
    Free registration
    License
    Not stated
    Data type
    PCAP, feature CSV
    Size
    Not stated
    Time span
    2023
    Link checked
    2026-09-27

    Dataset page for CICIoT2023Paper to cite for CICIoT2023

  • CICIoV2024

    IoT, ICS, and OT securityVehicle networks

    CICIoV2024 contains Controller Area Network (CAN) bus traffic captured from the electronic control units of a 2019 Ford vehicle. It includes benign messages plus DoS and spoofing attacks for Internet of Vehicles (IoV) research.

    Known weaknesses

    Data comes from one vehicle model, which limits generalization to other vehicles. The CAN message structure is simple, so detection scores may be optimistic.

    Access
    Free registration
    License
    Not stated
    Data type
    CAN message CSV
    Size
    Not stated
    Time span
    2024
    Link checked
    2026-09-27

    Dataset page for CICIoV2024Paper to cite for CICIoV2024

  • Edge-IIoTset

    IoT, ICS, and OT securityIoT and IIoT intrusion detection

    Edge-IIoTset contains traffic from an IoT and IIoT testbed with more than ten device types and fourteen attacks. It supports centralized and federated learning approaches to intrusion detection.

    Known weaknesses

    Attacks were generated in a lab and may not reflect production IIoT systems. The IEEE DataPort copy requires a paid subscription, so the authors' Kaggle copy is listed.

    Access
    Free registration
    License
    CC BY-NC-SA 4.0
    Data type
    PCAP, CSV
    Size
    1.48 GB zipped
    Time span
    2022
    Link checked
    2026-09-27

    Dataset page for Edge-IIoTsetPaper to cite for Edge-IIoTset

  • HAI (HIL-based Augmented ICS) Security Dataset

    IoT, ICS, and OT securityICS and SCADA

    HAI contains time-series process data from a hardware-in-the-loop (HIL) testbed combining turbine, boiler, and water treatment processes. Releases from 2020 to 2023 include normal operation and labeled attacks.

    Known weaknesses

    Versions differ in attacks, tags, and label columns, so results are not comparable across versions. Files from version 22.04 onward require Git LFS to download.

    Access
    Open download
    License
    CC BY-SA 4.0
    Data type
    Time-series CSV
    Size
    Not stated
    Time span
    2020-2023 releases
    Link checked
    2026-09-27

    Dataset page for HAI (HIL-based Augmented ICS) Security DatasetPaper to cite for HAI (HIL-based Augmented ICS) Security Dataset

  • IoT-23

    IoT, ICS, and OT securityIoT botnet traffic

    IoT-23 contains 20 malware captures from infected IoT devices and 3 captures of benign IoT devices. Each capture includes the PCAP and a Zeek connection log with malicious and benign labels.

    Known weaknesses

    A few scenarios contribute most of the flows, and port scan labels dominate. Only 3 of the 23 captures contain benign traffic from uninfected devices.

    Access
    Open download
    License
    CC BY 4.0
    Data type
    PCAP, labeled Zeek conn.log
    Size
    20 GB full, 8.8 GB labeled flows only
    Time span
    2018-2019
    Link checked
    2026-09-27

    Dataset page for IoT-23

  • Kitsune Network Attack Dataset

    IoT, ICS, and OT securityIoT intrusion detection

    The Kitsune dataset contains nine network captures of attacks on an IoT and video surveillance network. Each capture has per-packet features, labels, and the original packet capture.

    Known weaknesses

    Packets in the raw captures were truncated to 200 bytes for privacy. Each capture contains a single attack type, which simplifies the detection task.

    Access
    Open download
    License
    CC BY 4.0
    Data type
    PCAPNG, feature CSV
    Size
    17.7 GB
    Time span
    2018
    Link checked
    2026-09-27

    Dataset page for Kitsune Network Attack DatasetPaper to cite for Kitsune Network Attack Dataset

  • Mississippi State ICS Attack Datasets (Gas Pipeline and Power System)

    IoT, ICS, and OT securityICS and SCADA

    Tommy Morris and colleagues released SCADA (supervisory control and data acquisition) datasets from gas pipeline, water tank, and power testbeds. The data supports intrusion detection research for industrial control systems (ICS).

    Known weaknesses

    The authors warn that some datasets contain unintended patterns that make attacks trivially separable. The testbeds are small laboratory systems rather than operational plants.

    Access
    Open download
    License
    Not stated
    Data type
    ARFF, CSV, PCAP
    Size
    Not stated
    Time span
    2014-2015
    Link checked
    2026-09-27

    Dataset page for Mississippi State ICS Attack Datasets (Gas Pipeline and Power System)Paper to cite for Mississippi State ICS Attack Datasets (Gas Pipeline and Power System)

  • MQTT-IoT-IDS2020

    IoT, ICS, and OT securityMQTT and IoT protocols

    MQTT-IoT-IDS2020 contains traffic from a simulated network using MQTT (Message Queuing Telemetry Transport). It includes aggressive scan, UDP scan, SSH brute force, and MQTT brute force scenarios for IoT IDS research.

    Known weaknesses

    Sensors were simulated, and the attack set covers only four scenarios. Downloading the files requires a free account on the IEEE DataPort site.

    Access
    Free registration
    License
    Not stated
    Data type
    PCAP, packet and flow CSV
    Size
    1.35 GB PCAP
    Time span
    2020
    Link checked
    2026-09-27

    Dataset page for MQTT-IoT-IDS2020Paper to cite for MQTT-IoT-IDS2020

  • MQTTset

    IoT, ICS, and OT securityMQTT and IoT protocols

    MQTTset contains traffic from IoT sensors communicating through an MQTT broker, plus several MQTT-specific attacks. It supports machine learning detection of attacks on MQTT networks.

    Known weaknesses

    The sensor network was simulated with a fixed set of sensor types. Kaggle requires a free user account before files can be downloaded.

    Access
    Free registration
    License
    CC BY-NC-SA 4.0
    Data type
    PCAP, feature CSV
    Size
    Not stated
    Time span
    2020
    Link checked
    2026-09-27

    Dataset page for MQTTsetPaper to cite for MQTTset

  • N-BaIoT

    IoT, ICS, and OT securityIoT botnet traffic

    N-BaIoT contains traffic features from nine commercial IoT devices infected by Mirai and BASHLITE botnets. Researchers use it for per-device anomaly detection of IoT botnet attacks.

    Known weaknesses

    Only precomputed statistical features are released, so raw packets are unavailable. Published models often reach near-perfect accuracy, so the benchmark offers limited challenge.

    Access
    Open download
    License
    CC BY 4.0
    Data type
    Feature CSV
    Size
    1.7 GB
    Time span
    2018
    Link checked
    2026-09-27

    Dataset page for N-BaIoTPaper to cite for N-BaIoT

  • SWaT (Secure Water Treatment)

    IoT, ICS, and OT securityICS and SCADA

    SWaT contains sensor, actuator, and network data from a six-stage water treatment testbed at SUTD iTrust. The main release covers 51 sensors and actuators over 11 days, with 41 attacks on 4 days.

    Known weaknesses

    Version 0 of the normal data starts during a 30-minute tank emptying transient. Since July 2022, the iTrust team no longer answers questions about its datasets.

    Access
    Academic request
    License
    Free for research with credit to iTrust
    Data type
    Historian CSV, PCAP
    Size
    Not stated
    Time span
    December 2015, with later runs to 2023
    Link checked
    2026-09-27

    Dataset page for SWaT (Secure Water Treatment)Paper to cite for SWaT (Secure Water Treatment)

  • TON_IoT

    IoT, ICS, and OT securityIoT and IIoT intrusion detection

    TON_IoT combines IoT and IIoT (Industrial IoT) sensor telemetry, Windows and Linux host data, and network traffic from one testbed. Researchers use it for intrusion detection, threat hunting, and digital forensics.

    Known weaknesses

    Labels were assigned by tagging attacker IP addresses and timestamps. Booij et al. (2021) discuss heterogeneity of its features and attack types across the subsets.

    Access
    Open download
    License
    Free for academic research, commercial use requires author permission
    Data type
    PCAP, Zeek logs, CSV
    Size
    Not stated
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for TON_IoTPaper to cite for TON_IoT

  • UNSW IoT Traffic Traces

    IoT, ICS, and OT securityIoT device identification

    The UNSW IoT Analytics group released 20 daily traces from a smart environment testbed with consumer IoT devices. The data supports device identification and profiling from network traffic.

    Known weaknesses

    The traces contain no attack labels and suit device profiling rather than intrusion detection. Device firmware and cloud endpoints have changed since the 2016 capture period.

    Access
    Open download
    License
    Not stated
    Data type
    PCAP, CSV
    Size
    20 daily traces
    Time span
    September 23 to October 12, 2016
    Link checked
    2026-09-27

    Dataset page for UNSW IoT Traffic TracesPaper to cite for UNSW IoT Traffic Traces

  • WADI (Water Distribution)

    IoT, ICS, and OT securityICS and SCADA

    WADI contains 16 days of data from the iTrust water distribution testbed with 123 sensors and actuators. The release covers 14 normal days and 2 days with 15 attacks.

    Known weaknesses

    The plant was unstable during some periods, and iTrust removed affected readings in a revised file. Attack labels were added in a separate file after the first release.

    Access
    Academic request
    License
    Free for research with credit to iTrust
    Data type
    Historian CSV
    Size
    Not stated
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for WADI (Water Distribution)Paper to cite for WADI (Water Distribution)

  • WUSTL-EHMS-2020

    IoT, ICS, and OT securityMedical IoT

    WUSTL-EHMS-2020 combines network flow metrics and patient biometric readings from an enhanced healthcare monitoring system testbed. It supports intrusion detection using both network and medical data.

    Known weaknesses

    The dataset is small and comes from a single healthcare monitoring testbed. Labels were assigned by attacker MAC address, so address fields should be removed before training.

    Access
    Open download
    License
    Not stated
    Data type
    CSV
    Size
    3.9 MB
    Time span
    2020
    Link checked
    2026-09-27

    Dataset page for WUSTL-EHMS-2020Paper to cite for WUSTL-EHMS-2020

  • WUSTL-IIoT-2021

    IoT, ICS, and OT securityIoT and IIoT intrusion detection

    WUSTL-IIoT-2021 contains network flow data from an IIoT testbed that emulates industrial production systems. The attacks include command injection, reconnaissance, and DoS against the testbed.

    Known weaknesses

    About 92.7 percent of records are normal traffic, so classes are highly imbalanced. The authors advise dropping address and time columns that reveal attack labels.

    Access
    Open download
    License
    Not stated
    Data type
    Flow CSV
    Size
    Not stated
    Time span
    2021
    Link checked
    2026-09-27

    Dataset page for WUSTL-IIoT-2021

  • AndroZoo

    Malware and binariesAndroid app corpus

    AndroZoo is a large collection of Android Package (APK) files gathered from Google Play and other markets. Each app includes antivirus scan results that support malware and benign labeling.

    Known weaknesses

    API keys require a request from an institutional email and expire after six months. Downloads are capped at 500,000 APKs during each six-month key period. Contains live malware. Handle only in an isolated lab environment.

    Access
    Academic request
    License
    AndroZoo access conditions, citation required
    Data type
    Android APK files and metadata
    Size
    About 27.7 million APKs
    Time span
    Ongoing
    Link checked
    2026-09-27

    Dataset page for AndroZooPaper to cite for AndroZoo

  • Assemblage

    Malware and binariesBenign binary corpus

    Assemblage is a corpus of compiled binaries built from open-source C, C++, and Rust projects. Many binaries include debug information that links code back to source.

    Known weaknesses

    The corpus is benign only and needs pairing with malware for detection tasks. Build settings may differ from those used for real commercial software.

    Access
    Open download
    License
    MIT for tools, each binary keeps its source license
    Data type
    Windows PE and Linux ELF binaries with debug data
    Size
    About 567,000 binaries
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for AssemblagePaper to cite for Assemblage

  • Avast-CTU Public CAPEv2 Dataset

    Malware and binariesSandbox behavior reports

    This dataset from Avast and Czech Technical University (CTU) holds CAPEv2 sandbox reports for malicious Windows files. Reports cover ten malware families and six malware types, including banking Trojans and keyloggers.

    Known weaknesses

    Malware binaries are not included, so only reports are available. Two families, Emotet and Swisyn, make up over half of all samples.

    Access
    Open download
    License
    MIT
    Data type
    CAPEv2 JSON reports, full and reduced
    Size
    48,976 reports, 13 GB full or 566 MB reduced
    Time span
    July to September 2021
    Link checked
    2026-09-27

    Dataset page for Avast-CTU Public CAPEv2 DatasetPaper to cite for Avast-CTU Public CAPEv2 Dataset

  • BinKit

    Malware and binariesBenign binary similarity corpus

    BinKit is a benchmark for binary code similarity analysis built from GNU packages. It compiles each package across compilers, optimization levels, and architectures.

    Known weaknesses

    The binaries are all benign and target the Linux operating system only. Source packages are limited to GNU projects, which narrows code diversity.

    Access
    Open download
    License
    MIT
    Data type
    Compiled ELF binaries and extracted features
    Size
    371,928 binaries across 8 architectures
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for BinKitPaper to cite for BinKit

  • BODMAS

    Malware and binariesWindows PE malware with family labels

    BODMAS is a timestamped Windows Portable Executable (PE) dataset built for temporal analysis of malware classifiers. It provides EMBER-compatible features and family labels for malicious samples.

    Known weaknesses

    Features and metadata are public, but malware binaries require an email request with affiliation checks. Benign binaries are not shared with requesters because of software copyright concerns. Contains live malware. Handle only in an isolated lab environment.

    Access
    Academic request
    License
    Research use with citation, no redistribution
    Data type
    Feature vectors, metadata, malware binaries on request
    Size
    57,293 malware and 77,142 benign samples across 581 families
    Time span
    August 2019 to September 2020
    Link checked
    2026-09-27

    Dataset page for BODMASPaper to cite for BODMAS

  • CCCS-CIC-AndMal-2020

    Malware and binariesAndroid malware static and dynamic features

    This dataset from the Canadian Centre for Cyber Security (CCCS) and CIC covers 14 Android malware categories. It provides static manifest features and dynamic features from memory, application programming interface (API) calls, network, and logs.

    Known weaknesses

    Labels come from VirusTotal detections and may contain some labeling noise. Some malware families have very few samples, which limits family-level classification.

    Access
    Open download
    License
    Redistribution allowed with citation
    Data type
    Static and dynamic feature CSVs
    Size
    400,000 apps, half benign and half malware, 191 families
    Time span
    2020
    Link checked
    2026-09-27

    Dataset page for CCCS-CIC-AndMal-2020

  • CIC-AndMal2017

    Malware and binariesAndroid malware network behavior

    CIC-AndMal2017 captures network traffic from Android malware and benign apps run on real devices. It covers adware, ransomware, scareware, and Short Message Service (SMS) malware across 42 families.

    Known weaknesses

    Flow features alone give weak family separation in later studies. The companion CIC-InvesAndMal2019 release adds API call sequences and device logs. Contains live malware. Handle only in an isolated lab environment.

    Access
    Open download
    License
    Free for research with citation
    Data type
    PCAP files and flow features, APKs
    Size
    10,854 apps with 4,354 malware and 6,500 benign
    Time span
    2015 to 2017
    Link checked
    2026-09-27

    Dataset page for CIC-AndMal2017Paper to cite for CIC-AndMal2017

  • CIC-Evasive-PDFMal2022

    Malware and binariesMalicious PDF documents

    CIC-Evasive-PDFMal2022 targets detection of evasive malicious Portable Document Format (PDF) files. It provides general and structural features for benign and malicious documents.

    Known weaknesses

    The dataset contains 37 static features only and no raw PDF files. Evasive samples were chosen to reduce overlap with older PDF datasets.

    Access
    Open download
    License
    Redistribution allowed with citation
    Data type
    Static feature CSV
    Size
    10,025 records with 5,557 malicious and 4,468 benign
    Time span
    2022
    Link checked
    2026-09-27

    Dataset page for CIC-Evasive-PDFMal2022

  • CIC-MalMem-2022

    Malware and binariesObfuscated malware memory analysis

    CIC-MalMem-2022 targets detection of obfuscated Windows malware through analysis of memory dumps. It balances benign records with spyware, ransomware, and Trojan families.

    Known weaknesses

    The release uses features extracted from memory dumps rather than full dumps. Each malicious category contains only five families, which limits family diversity.

    Access
    Open download
    License
    Free for research with citation
    Data type
    Memory-derived feature records (CSV)
    Size
    58,596 records, half benign and half malicious
    Time span
    2022
    Link checked
    2026-09-27

    Dataset page for CIC-MalMem-2022Paper to cite for CIC-MalMem-2022

  • CICMalDroid 2020

    Malware and binariesAndroid malware categories

    CICMalDroid 2020 from the Canadian Institute for Cybersecurity (CIC) covers four Android malware categories and benign apps. It includes Android Package (APK) files, dynamic analysis logs, and static and dynamic feature files.

    Known weaknesses

    Contains live malware. Handle only in an isolated lab environment. Not every APK completed dynamic analysis, so feature files cover a subset.

    Access
    Open download
    License
    Free for research with citation
    Data type
    APKs, dynamic logs (JSON), feature CSVs, PCAP files
    Size
    About 17,341 APKs, 11,598 in the feature sets
    Time span
    December 2017 to December 2018
    Link checked
    2026-09-27

    Dataset page for CICMalDroid 2020Paper to cite for CICMalDroid 2020

  • DikeDataset

    Malware and binariesLabeled PE and OLE files

    DikeDataset contains benign and malicious Portable Executable (PE) and Object Linking and Embedding (OLE) files. Each file has a malice score and family membership scores.

    Known weaknesses

    Contains live malware. Handle only in an isolated lab environment. Antivirus software may flag, quarantine, or delete files during download.

    Access
    Open download
    License
    MIT
    Data type
    PE and OLE binaries with labels
    Size
    Not stated
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for DikeDataset

  • Drebin

    Malware and binariesAndroid malware samples

    Drebin is a classic Android malware dataset collected for explainable malware detection research. It groups the malicious Android samples into named malware families for classification.

    Known weaknesses

    The samples are more than ten years old and target outdated Android versions. Access requires an email request with identity and advisor verification. Contains live malware. Handle only in an isolated lab environment.

    Access
    Academic request
    License
    No redistribution, citation required
    Data type
    Android APK malware samples and features
    Size
    5,560 apps from 179 families
    Time span
    August 2010 to October 2012
    Link checked
    2026-09-27

    Dataset page for DrebinPaper to cite for Drebin

  • EMBER 2017 and 2018

    Malware and binariesWindows PE malware features

    The Endgame Malware BEnchmark for Research (EMBER) provides static features from Windows Portable Executable (PE) files. Each release pairs extracted features with benign, malicious, or unlabeled tags for supervised learning.

    Known weaknesses

    The public release contains extracted features only and includes no raw binary files. The two releases used different sample selection criteria and feature versions. The repository was archived in 2026 and is now read-only.

    Access
    Open download
    License
    MIT for data files, AGPL-3.0 for code
    Data type
    Extracted static features (JSON) with labels
    Size
    About 1.1 million PE files (2017) and 1 million PE files (2018)
    Time span
    Samples first seen in or before 2017 and 2018
    Link checked
    2026-09-27

    Dataset page for EMBER 2017 and 2018Paper to cite for EMBER 2017 and 2018

  • EMBER2024

    Malware and binariesMulti-platform malware features

    EMBER2024 is a benchmark of features and labels for files seen on VirusTotal between 2023 and 2024. It covers six file formats and supports detection, family, behavior, packer, and threat group tasks.

    Known weaknesses

    The release contains features and labels but no raw binaries. Labels derive from antivirus detections and may carry some noise.

    Access
    Open download
    License
    Apache-2.0
    Data type
    Extracted features, labels, and baseline models
    Size
    About 3.2 million files across Win32, Win64, .NET, APK, ELF, and PDF
    Time span
    September 2023 to December 2024
    Link checked
    2026-09-27

    Dataset page for EMBER2024Paper to cite for EMBER2024

  • IoTPOT Malware Dataset

    Malware and binariesIoT and Linux ELF malware

    IoTPOT is a honeypot from Yokohama National University that captures Internet of Things (IoT) malware. Shared data include captured Linux Executable and Linkable Format (ELF) binaries and the malware download URLs observed.

    Known weaknesses

    Contains live malware. Handle only in an isolated lab environment. Access requires a data request form that the operators review before sharing. Samples are dominated by Mirai and Gafgyt variants that target weak IoT credentials.

    Access
    Academic request
    License
    Not stated
    Data type
    ELF malware binaries and download URLs
    Size
    About 16,000 files for 2024 plus X-Pot captures
    Time span
    2015 to present
    Link checked
    2026-09-27

    Dataset page for IoTPOT Malware DatasetPaper to cite for IoTPOT Malware Dataset

  • KronoDroid

    Malware and binariesAndroid malware with time labels

    KronoDroid is a hybrid Android dataset with static and dynamic features and timestamps for each app. It includes separate captures from an emulator and from real devices.

    Known weaknesses

    Feature CSVs are public, but raw logs and Android Package (APK) files require researcher verification. The repository has no formal license file beyond the citation request. Contains live malware. Handle only in an isolated lab environment.

    Access
    Open download
    License
    Research use with citation
    Data type
    Feature CSVs, raw logs and APKs on request
    Size
    About 78,000 real-device and 64,000 emulator samples, 240 families
    Time span
    2008 to 2020
    Link checked
    2026-09-27

    Dataset page for KronoDroidPaper to cite for KronoDroid

  • Mal-API-2019

    Malware and binariesMalware API call sequences

    Mal-API-2019 contains Windows application programming interface (API) call sequences recorded from malware in the Cuckoo sandbox. Each sequence carries one of eight malware type labels such as Trojan or worm.

    Known weaknesses

    The dataset has no benign samples, so it cannot support detection tasks alone. Class labels come from antivirus names and are coarse malware type categories.

    Access
    Open download
    License
    MIT
    Data type
    API call sequences (text)
    Size
    7,107 samples in 8 classes
    Time span
    2019
    Link checked
    2026-09-27

    Dataset page for Mal-API-2019Paper to cite for Mal-API-2019

  • MalDICT

    Malware and binariesMalware behaviors, platforms, exploits, and packers

    MalDICT provides four tagged malware benchmarks covering behaviors, target platforms, exploited vulnerabilities, and packers. Tags come from antivirus results and use a temporal train and test split.

    Known weaknesses

    Only Portable Executable (PE) binaries are published, and other files must be fetched from VirusShare by hash. Tags derive from antivirus labels and therefore carry some amount of label noise. Contains live malware. Handle only in an isolated lab environment.

    Access
    Open download
    License
    Apache-2.0
    Data type
    File hashes, tags, EMBER features, disarmed PE binaries
    Size
    About 4.3 million files in the behavior set plus smaller platform, vulnerability, and packer sets
    Time span
    VirusShare chunks 0 to 465
    Link checked
    2026-09-27

    Dataset page for MalDICTPaper to cite for MalDICT

  • MaleVis

    Malware and binariesMalware images

    MaleVis provides RGB byte images of Windows malware and legitimate software for vision research. Images come in two resolutions for convolutional neural network (CNN) experiments.

    Known weaknesses

    Access requires a short Google Form registration before the download link appears. Only images are shared, and binaries require direct contact with the authors.

    Access
    Free registration
    License
    Academic use only
    Data type
    RGB images
    Size
    14,226 images in 26 classes
    Time span
    2019
    Link checked
    2026-09-27

    Dataset page for MaleVis

  • MaleX

    Malware and binariesMalware images and features

    MaleX is a curated set of Windows executables with malware and benign labels. It releases frequency-domain images and extracted features for image-based malware detection.

    Known weaknesses

    The underlying executable binaries are not currently available for download. Access requires a form that asks for affiliation and intended research use.

    Access
    Free registration
    License
    GPL-3.0
    Data type
    Images and feature files
    Size
    1,044,394 executables with 864,669 malware and 179,725 benign
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for MaleXPaper to cite for MaleX

  • Malimg

    Malware and binariesMalware images

    Malimg converts Windows malware binaries into grayscale byteplot images for visual classification. It is a common baseline dataset for image-based malware classification research.

    Known weaknesses

    The download link comes from the authors' company page in the MaleX repository. Many studies report near-perfect accuracy, so the benchmark is largely saturated. The underlying samples date from 2011 and do not reflect current malware.

    Access
    Open download
    License
    Not stated
    Data type
    Grayscale PNG images
    Size
    9,339 images from 25 families
    Time span
    2011
    Link checked
    2026-09-27

    Dataset page for MalimgPaper to cite for Malimg

  • MalNet

    Malware and binariesMalware graphs and images

    MalNet provides function call graphs and binary images derived from Android Package (APK) files in AndroZoo. It supports graph learning and image-based malware type and family classification.

    Known weaknesses

    Labels come from antivirus detections and therefore carry some labeling noise. Class distribution is highly imbalanced across the malware types and families.

    Access
    Open download
    License
    Not stated
    Data type
    Function call graphs and byteplot images
    Size
    About 1.2 million graphs and 1.2 million images across 47 types and 696 families
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for MalNetPaper to cite for MalNet

  • MalShare

    Malware and binariesLive malware sample repository

    MalShare is a community malware repository that has operated since 2013. Registered users can search, upload, and download samples through the site or its API.

    Known weaknesses

    Contains live malware. Handle only in an isolated lab environment. Samples are user submissions and carry only limited labels or family information.

    Access
    Free registration
    License
    Terms of service on the site
    Data type
    Malware samples and daily hash lists
    Size
    Not stated
    Time span
    2013 to present
    Link checked
    2026-09-27

    Dataset page for MalShare

  • MalwareBazaar

    Malware and binariesLive malware sample repository

    MalwareBazaar is an abuse.ch project that collects and shares malware samples with tags and signatures. Users can query samples and download them through the web interface or an application programming interface (API).

    Known weaknesses

    Contains live malware. Handle only in an isolated lab environment. Programmatic API use requires a free authentication key from the abuse.ch portal. Samples reflect community submissions rather than a balanced research sample.

    Access
    API
    License
    Terms of service on the site
    Data type
    Malware samples, hashes, tags, and vendor intelligence
    Size
    Not stated
    Time span
    Ongoing
    Link checked
    2026-09-27

    Dataset page for MalwareBazaar

  • MH-100K

    Malware and binariesAndroid malware features

    MH-100K is an Android dataset with permissions, application programming interface (API) calls, intents, and VirusTotal results for each app. All apps were drawn from the AndroZoo repository and scanned with VirusTotal.

    Known weaknesses

    Malware labels depend on a threshold of four VirusTotal detections. The class balance is heavily skewed toward benign apps over malicious ones.

    Access
    Open download
    License
    CC BY 4.0
    Data type
    Feature matrix and VirusTotal metadata
    Size
    101,975 apps with 9,800 labeled malicious
    Time span
    2010 to 2022
    Link checked
    2026-09-27

    Dataset page for MH-100KPaper to cite for MH-100K

  • Microsoft Malware Classification Challenge (BIG 2015)

    Malware and binariesWindows malware family classification

    This Kaggle challenge released Windows malware as hexadecimal byte dumps and disassembly files. Researchers use it as a standard benchmark for malware family classification.

    Known weaknesses

    Headers were removed, so the files cannot run and header analysis is not possible. Labels for the Kaggle test set were never released to the public. The samples date from 2015 and do not reflect current threats.

    Access
    Free registration
    License
    Kaggle competition rules
    Data type
    Hex byte dumps and IDA disassembly (.asm) files
    Size
    About 21,700 samples in 9 families, nearly 0.5 TB uncompressed
    Time span
    2015
    Link checked
    2026-09-27

    Dataset page for Microsoft Malware Classification Challenge (BIG 2015)Paper to cite for Microsoft Malware Classification Challenge (BIG 2015)

  • Microsoft Malware Prediction

    Malware and binariesEndpoint infection prediction

    This Kaggle competition dataset predicts whether a Windows machine will encounter malware. Each row describes machine properties from Windows Defender heartbeat and threat telemetry.

    Known weaknesses

    The data describes machine configurations and does not include any malware files. The training and test sets differ in time, which causes distribution shift.

    Access
    Free registration
    License
    Kaggle competition rules
    Data type
    Tabular endpoint telemetry (CSV)
    Size
    About 8.9 million training rows with 83 features
    Time span
    2018 telemetry, released 2019
    Link checked
    2026-09-27

    Dataset page for Microsoft Malware Prediction

  • MOTIF

    Malware and binariesMalware family ground truth

    The Malware Open-source Threat Intelligence Family (MOTIF) dataset links Portable Executable (PE) malware to families named in public threat reports. It offers disarmed binaries and EMBER features with report-based family labels.

    Known weaknesses

    Samples are disarmed to block execution but still contain real malicious code. Many families have only a few samples, which limits per-family training. Contains live malware. Handle only in an isolated lab environment.

    Access
    Open download
    License
    Booz Allen Public License v1.0
    Data type
    Disarmed PE binaries, EMBER v2 features, family labels
    Size
    3,095 samples from 454 families, about 1.47 GB compressed
    Time span
    Threat reports from 2016 to 2020
    Link checked
    2026-09-27

    Dataset page for MOTIFPaper to cite for MOTIF

  • RanSAP

    Malware and binariesRansomware storage access patterns

    RanSAP records low-level storage access patterns from ransomware and benign programs on Windows. Data were collected through a hypervisor on hard disk and solid state drive setups.

    Known weaknesses

    The sample count is small and uses mainly Windows 7 hosts. The no-derivatives license limits sharing of modified or derived versions of the data.

    Access
    Open download
    License
    CC BY-ND 4.0
    Data type
    Storage read and write traces with entropy
    Size
    7 ransomware samples, 21 extra variants, 5 benign programs
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for RanSAPPaper to cite for RanSAP

  • RISS Ransomware Dataset

    Malware and binariesRansomware dynamic behavior

    This dataset from the Research Institute in Information Security (RISS) records early dynamic behavior of ransomware and goodware. Features come from Cuckoo sandbox runs that monitor the first phases of execution.

    Known weaknesses

    The samples date from 2016 and miss modern ransomware families. Only extracted features are shared, and the original binaries are not included.

    Access
    Open download
    License
    Not stated
    Data type
    Dynamic analysis feature matrix
    Size
    582 ransomware and 942 goodware samples
    Time span
    2016
    Link checked
    2026-09-27

    Dataset page for RISS Ransomware DatasetPaper to cite for RISS Ransomware Dataset

  • SOREL-20M

    Malware and binariesWindows PE malware samples and features

    The Sophos-ReversingLabs 20 Million dataset (SOREL-20M) supports research on malicious Portable Executable (PE) file detection. It offers labels, EMBER features, and disarmed binaries on Amazon Simple Storage Service (S3).

    Known weaknesses

    Binaries are disarmed by zeroing header fields but still contain real malicious code. Benign binaries are not distributed because of intellectual property limits. The full download is very large and needs substantial storage. Contains live malware. Handle only in an isolated lab environment.

    Access
    Open download
    License
    Apache-2.0
    Data type
    Disarmed PE binaries, EMBER v2 features, PE metadata
    Size
    About 20 million samples, roughly 8 TB with binaries
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for SOREL-20MPaper to cite for SOREL-20M

  • theZoo

    Malware and binariesLive malware sample repository

    theZoo is a public GitHub repository of live malware binaries and source code for education. Each sample ships in a password-protected archive with matching hash and password files.

    Known weaknesses

    Contains live malware. Handle only in an isolated lab environment. Some samples are worms that can spread automatically across connected networks. Coverage is curated and uneven, so it suits teaching more than training models.

    Access
    Open download
    License
    GPL-3.0 for code, which excludes the malware samples
    Data type
    Encrypted malware binaries and source code
    Size
    Not stated
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for theZoo

  • VirusShare

    Malware and binariesLive malware sample repository

    VirusShare is a long-running repository of malware samples for researchers and incident responders. Hash lists and sample archives support large-scale malware corpus construction.

    Known weaknesses

    Contains live malware. Handle only in an isolated lab environment. Accounts are granted by invitation after a request to the operator. Samples come without family labels and without any benign counterparts for comparison.

    Access
    Free registration
    License
    Not stated
    Data type
    Malware samples and hash lists
    Size
    More than 119 million samples
    Time span
    Ongoing
    Link checked
    2026-09-27

    Dataset page for VirusShare

  • 5G-NIDD

    Network traffic and intrusion detectionWireless and 5G

    5G-NIDD contains traffic from a 5G test network at the University of Oulu with two base stations. Attacker nodes launch DoS floods and port scans against a server in the mobile edge environment.

    Known weaknesses

    The attack set is limited to DoS floods and port scans. Downloading the files requires a free account on the IEEE DataPort site.

    Access
    Free registration
    License
    Not stated
    Data type
    PCAPNG, Argus, CSV
    Size
    About 3.6 GB PCAP across two base stations
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for 5G-NIDDPaper to cite for 5G-NIDD

  • AWID3

    Network traffic and intrusion detectionWireless and 5G

    AWID3 contains IEEE 802.11 enterprise wireless traffic with attacks such as deauthentication, evil twin, and website spoofing. It also includes common wired-network attacks and a CSV version with 254 features.

    Known weaknesses

    Data comes from a small lab testbed with a limited set of devices. Access requires submitting a download request form to the dataset authors.

    Access
    Academic request
    License
    Not stated
    Data type
    PCAP, feature CSV
    Size
    Not stated
    Time span
    2021
    Link checked
    2026-09-27

    Dataset page for AWID3Paper to cite for AWID3

  • CAIDA Anonymized Internet Traces

    Network traffic and intrusion detectionInternet measurement

    CAIDA collected anonymized packet header traces from high-speed commercial backbone links in the United States. Researchers use them to study internet traffic characteristics and build realistic baselines.

    Known weaknesses

    Packet payloads are removed and IP addresses are anonymized before release. Collection ended in January 2019, so the traces do not reflect current traffic.

    Access
    Academic request
    License
    CAIDA Acceptable Use Agreement
    Data type
    PCAP (headers only)
    Size
    Not stated
    Time span
    April 2008 to January 2019
    Link checked
    2026-09-27

    Dataset page for CAIDA Anonymized Internet Traces

  • CAIDA DDoS Attack 2007

    Network traffic and intrusion detectionDDoS

    The Center for Applied Internet Data Analysis (CAIDA) released one hour of anonymized traffic from a real DDoS attack. The trace holds only attack traffic to the victim and its responses.

    Known weaknesses

    The trace is old and covers a single attack event. Non-attack traffic was removed, so it cannot train a benign baseline alone.

    Access
    Academic request
    License
    CAIDA Acceptable Use Agreement
    Data type
    PCAP
    Size
    5.3 GB compressed, 21 GB uncompressed
    Time span
    August 4, 2007
    Link checked
    2026-09-27

    Dataset page for CAIDA DDoS Attack 2007

  • CESNET-QUIC22

    Network traffic and intrusion detectionEncrypted traffic classification

    CESNET-QUIC22 contains one month of QUIC (Quick UDP Internet Connections) flows from Czech national backbone links. Labels identify web services, which supports classification of encrypted QUIC traffic.

    Known weaknesses

    Service labels are derived from server names seen in handshakes and may contain errors. The dataset has no attack labels and suits traffic classification rather than intrusion detection.

    Access
    Open download
    License
    CC BY 4.0
    Data type
    Flow CSV with packet sequences
    Size
    21 GB
    Time span
    2022
    Link checked
    2026-09-27

    Dataset page for CESNET-QUIC22Paper to cite for CESNET-QUIC22

  • CESNET-TLS-Year22

    Network traffic and intrusion detectionEncrypted traffic classification

    CESNET-TLS-Year22 contains a full year of TLS (Transport Layer Security) flows from Czech backbone lines. It supports encrypted traffic classification and studies of model drift over time.

    Known weaknesses

    Labels identify services rather than malicious behavior, so the data does not support attack detection directly. The CESNET DataZoo Python library is recommended for handling the data volume.

    Access
    Open download
    License
    CC BY 4.0
    Data type
    Flow data with packet sequences
    Size
    30.5 GB
    Time span
    2022
    Link checked
    2026-09-27

    Dataset page for CESNET-TLS-Year22Paper to cite for CESNET-TLS-Year22

  • CIC-Bell-DNS 2021

    Network traffic and intrusion detectionDNS and DNS-over-HTTPS

    CIC-Bell-DNS 2021 contains DNS features for 400,000 benign and 13,011 malicious domain samples. Labels separate spam, phishing, and malware domains for malicious domain classification research.

    Known weaknesses

    Domain labels come from public blocklists that may contain stale or wrong entries. Benign samples far outnumber malicious samples, which creates strong class imbalance.

    Access
    Free registration
    License
    Redistribution allowed with citation
    Data type
    PCAP, feature CSV
    Size
    413,011 samples
    Time span
    2021
    Link checked
    2026-09-27

    Dataset page for CIC-Bell-DNS 2021Paper to cite for CIC-Bell-DNS 2021

  • CIC-Bell-DNS-EXF-2021

    Network traffic and intrusion detectionDNS and DNS-over-HTTPS

    CIC-Bell-DNS-EXF-2021 contains DNS traffic generated by exfiltrating files of varied sizes through DNS queries. It provides 30 stateless and stateful features for detecting DNS data exfiltration.

    Known weaknesses

    Exfiltration traffic was produced with the DNSExfiltrator tool using fixed encoding and throttling settings. Real attackers may use different encodings, tools, and timing patterns.

    Access
    Free registration
    License
    Redistribution allowed with citation
    Data type
    PCAP, feature CSV
    Size
    270.8 MB DNS traffic
    Time span
    2021
    Link checked
    2026-09-27

    Dataset page for CIC-Bell-DNS-EXF-2021Paper to cite for CIC-Bell-DNS-EXF-2021

  • CIC-Darknet2020

    Network traffic and intrusion detectionEncrypted traffic classification

    CIC-Darknet2020 combines ISCXTor2016 and ISCXVPN2016 into one darknet traffic dataset with two label layers. It supports detecting Tor and VPN traffic and classifying the underlying application type.

    Known weaknesses

    The dataset reuses 2016 captures and inherits their lab generation limits. Here darknet means Tor and VPN traffic, not network telescope data.

    Access
    Free registration
    License
    Redistribution allowed with citation
    Data type
    Flow CSV
    Size
    Not stated
    Time span
    2020
    Link checked
    2026-09-27

    Dataset page for CIC-Darknet2020Paper to cite for CIC-Darknet2020

  • CIC-DDoS2019

    Network traffic and intrusion detectionDDoS

    CIC-DDoS2019 contains benign traffic and 12 reflection and exploitation DDoS attacks over two capture days. Attacks include NTP, DNS, LDAP, MSSQL, NetBIOS, SNMP, SSDP, UDP, and SYN floods.

    Known weaknesses

    Flows were generated with CICFlowMeter-V3, the tool linked to documented labeling errors in CIC-IDS2017. Attack flows vastly outnumber benign flows in the released CSV files.

    Access
    Free registration
    License
    Redistribution allowed with citation
    Data type
    PCAP, flow CSV, event logs
    Size
    Not stated
    Time span
    2019
    Link checked
    2026-09-27

    Dataset page for CIC-DDoS2019Paper to cite for CIC-DDoS2019

  • CIC-IDS2017

    Network traffic and intrusion detectionNetwork intrusion detection

    CIC-IDS2017 holds five days of emulated enterprise traffic with benign activity and common network attacks. It provides packet capture (PCAP) files and labeled flow features for training intrusion detection systems (IDS).

    Known weaknesses

    Engelen et al. (2021) documented flow construction and labeling errors in the released CSV files. Benign traffic comes from scripted user profiles rather than real users. Attack classes such as Heartbleed and infiltration have very few samples.

    Access
    Free registration
    License
    Public for researchers with citation required
    Data type
    PCAP, flow CSV
    Size
    Not stated
    Time span
    July 3-7, 2017
    Link checked
    2026-09-27

    Dataset page for CIC-IDS2017Paper to cite for CIC-IDS2017

  • CIDDS-001 and CIDDS-002

    Network traffic and intrusion detectionNetwork intrusion detection

    The Coburg Intrusion Detection Data Sets (CIDDS) contain labeled NetFlow data from an emulated small business network in OpenStack. Scripts generate benign user behavior, and the attacks include DoS, brute force, and port scans.

    Known weaknesses

    Benign traffic is emulated by Python scripts rather than real users. Public IP addresses are anonymized, and only flow data is released.

    Access
    Open download
    License
    Not stated
    Data type
    NetFlow CSV
    Size
    384 MB (CIDDS-001) and 204 MB (CIDDS-002) zipped
    Time span
    2017
    Link checked
    2026-09-27

    Dataset page for CIDDS-001 and CIDDS-002

  • CIRA-CIC-DoHBrw-2020

    Network traffic and intrusion detectionDNS and DNS-over-HTTPS

    CIRA-CIC-DoHBrw-2020 captures benign DNS-over-HTTPS (DoH) traffic, malicious DoH tunnels, and non-DoH HTTPS browsing. It supports two-layer classification of encrypted DNS traffic with statistical and time-series features.

    Known weaknesses

    Malicious DoH traffic comes from three tunneling tools in a lab setting. Models may learn tool signatures rather than general tunnel behavior.

    Access
    Free registration
    License
    Redistribution allowed with citation
    Data type
    PCAP, flow CSV
    Size
    Not stated
    Time span
    2020
    Link checked
    2026-09-27

    Dataset page for CIRA-CIC-DoHBrw-2020Paper to cite for CIRA-CIC-DoHBrw-2020

  • CSE-CIC-IDS2018

    Network traffic and intrusion detectionNetwork intrusion detection

    CSE-CIC-IDS2018 contains traffic and host logs from a simulated organization with 420 machines and 30 servers. It covers seven attack scenarios, including brute force, botnet, DoS (denial of service), and web attacks, for IDS benchmarking.

    Known weaknesses

    Liu et al. (2022) reported labeling and flow extraction errors similar to those found in CIC-IDS2017. Download requires the AWS command line tool and substantial storage.

    Access
    Open download
    License
    Redistribution allowed with citation and link to the AWS page
    Data type
    PCAP, flow CSV, system logs
    Size
    Not stated
    Time span
    February 14 to March 2, 2018
    Link checked
    2026-09-27

    Dataset page for CSE-CIC-IDS2018Paper to cite for CSE-CIC-IDS2018

  • CTU-13

    Network traffic and intrusion detectionBotnet traffic

    CTU-13 contains 13 scenarios of real botnet traffic mixed with normal and background traffic at CTU University. Each scenario provides labeled bidirectional NetFlows and botnet-only packet captures.

    Known weaknesses

    The botnet families and network conditions date from 2011 and are now dated. Full PCAPs with background traffic are withheld for privacy reasons. Background flows carry no verified benign or malicious label in any scenario.

    Access
    Open download
    License
    Not stated
    Data type
    Labeled NetFlow, botnet PCAP
    Size
    1.9 GB compressed archive
    Time span
    2011
    Link checked
    2026-09-27

    Dataset page for CTU-13Paper to cite for CTU-13

  • CyberLab Honeynet Dataset

    Network traffic and intrusion detectionHoneypot data

    The CyberLab dataset contains logs from Cowrie SSH and Telnet honeypots on about 50 nodes. The logs record attacker sessions, commands, and downloaded files for threat research.

    Known weaknesses

    Public IP addresses are pseudonymized to protect the honeypot nodes. The number of active nodes changed during the experiment, which affects volume comparisons over time.

    Access
    Open download
    License
    CC BY 4.0
    Data type
    Cowrie JSON logs
    Size
    10 GB
    Time span
    May 2019 to February 2020
    Link checked
    2026-09-27

    Dataset page for CyberLab Honeynet Dataset

  • DARPA 1998 and 1999 Intrusion Detection Evaluation

    Network traffic and intrusion detectionNetwork intrusion detection (legacy)

    MIT Lincoln Laboratory released the DARPA (Defense Advanced Research Projects Agency) evaluation data with simulated network traffic and audit logs. It contains weeks of labeled attacks and attack-free weeks for training anomaly detectors.

    Known weaknesses

    This dataset is outdated and reflects network services and attacks from the late 1990s. McHugh (2000) criticized its synthetic background traffic and simulation artifacts. The 1999 data sits on a companion Lincoln Laboratory page.

    Access
    Open download
    License
    Not stated
    Data type
    PCAP (tcpdump), BSM audit logs
    Size
    Not stated
    Time span
    1998-1999
    Link checked
    2026-09-27

    Dataset page for DARPA 1998 and 1999 Intrusion Detection EvaluationPaper to cite for DARPA 1998 and 1999 Intrusion Detection Evaluation

  • DARPA Operationally Transparent Cyber (OpTC)

    Network traffic and intrusion detectionEnterprise network and authentication

    OpTC contains endpoint telemetry and Zeek network logs from a DARPA pilot that scaled monitoring to about one thousand hosts. It includes a red team scenario for studying attack detection across hosts.

    Known weaknesses

    The data is hosted on Google Drive and is very large. DARPA released it as is, without support or accuracy warranties.

    Access
    Open download
    License
    Public domain
    Data type
    eCAR JSON, Zeek logs
    Size
    Not stated
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for DARPA Operationally Transparent Cyber (OpTC)

  • HIKARI-2021

    Network traffic and intrusion detectionEncrypted traffic intrusion detection

    HIKARI-2021 combines real benign traffic with synthetic attacks carried over encrypted HTTPS sessions. It targets intrusion detection when payloads are encrypted and only flow features are usable.

    Known weaknesses

    Attack traffic is synthetic and generated in a lab setting. Attack classes are limited to brute force, probing, and cryptomining traffic.

    Access
    Open download
    License
    CC BY 4.0
    Data type
    PCAP, flow CSV
    Size
    11.7 GB
    Time span
    2021
    Link checked
    2026-09-27

    Dataset page for HIKARI-2021Paper to cite for HIKARI-2021

  • ISCX Botnet 2014

    Network traffic and intrusion detectionBotnet traffic

    ISCX Botnet 2014 merges traces from ISOT, ISCX-IDS2012, and the Malware Capture Facility Project into training and test sets. It supports evaluating botnet detection with 7 training and 16 test botnet types.

    Known weaknesses

    The dataset was built with an overlay method that maps botnet IP addresses into merged traces. Merging traces from different sources can leave artifacts that make detection easier.

    Access
    Free registration
    License
    Not stated
    Data type
    PCAP
    Size
    5.3 GB training, 8.5 GB test
    Time span
    2014
    Link checked
    2026-09-27

    Dataset page for ISCX Botnet 2014Paper to cite for ISCX Botnet 2014

  • ISCX Tor-nonTor 2016

    Network traffic and intrusion detectionEncrypted traffic classification

    ISCXTor2016 contains Tor and regular traffic for browsing, email, chat, streaming, file transfer, VoIP, and P2P activities. It supports detecting Tor use and classifying applications inside Tor traffic.

    Known weaknesses

    The traffic reflects Tor and application versions that were current in 2016. Lab generation by a few users limits the diversity of behavior in the data.

    Access
    Free registration
    License
    Public for researchers with citation required
    Data type
    PCAP, feature CSV
    Size
    22 GB
    Time span
    2016
    Link checked
    2026-09-27

    Dataset page for ISCX Tor-nonTor 2016Paper to cite for ISCX Tor-nonTor 2016

  • ISCX VPN-nonVPN 2016

    Network traffic and intrusion detectionEncrypted traffic classification

    ISCXVPN2016 captures regular and VPN (virtual private network) sessions across 14 traffic categories such as VoIP and streaming. It supports classification of encrypted and VPN traffic using time-related flow features.

    Known weaknesses

    Traffic was generated by a small set of lab users and applications in 2016. Application versions and VPN protocols have changed considerably since the capture period.

    Access
    Free registration
    License
    Public for researchers with citation required
    Data type
    PCAP, feature CSV
    Size
    28 GB
    Time span
    2016
    Link checked
    2026-09-27

    Dataset page for ISCX VPN-nonVPN 2016Paper to cite for ISCX VPN-nonVPN 2016

  • ISCX-IDS2012

    Network traffic and intrusion detectionNetwork intrusion detection

    ISCX-IDS2012 records seven days of testbed network activity at the University of New Brunswick. Attacks include infiltration, HTTP denial of service, IRC botnet DDoS (distributed denial of service), and SSH brute force.

    Known weaknesses

    The traffic dates from 2010 and lacks modern protocols and attack types. Each day contains a single attack scenario, which simplifies detection.

    Access
    Free registration
    License
    Not stated
    Data type
    PCAP, labeled flow XML
    Size
    About 84 GB PCAP across seven days
    Time span
    June 11-17, 2010
    Link checked
    2026-09-27

    Dataset page for ISCX-IDS2012Paper to cite for ISCX-IDS2012

  • Kyoto 2006+

    Network traffic and intrusion detectionHoneypot data

    Kyoto 2006+ contains session records from honeypots and servers deployed at Kyoto University. It provides traffic features and IDS alert labels for network intrusion detection research.

    Known weaknesses

    Labels depend on IDS and antivirus alerts, which miss unknown attacks. Traffic that reaches honeypots is mostly malicious by design, which differs from enterprise networks.

    Access
    Open download
    License
    Not stated
    Data type
    Session feature text files
    Size
    Not stated
    Time span
    November 2006 to December 2015
    Link checked
    2026-09-27

    Dataset page for Kyoto 2006+Paper to cite for Kyoto 2006+

  • LANL Comprehensive, Multi-Source Cyber-Security Events (2015)

    Network traffic and intrusion detectionEnterprise network and authentication

    Los Alamos National Laboratory (LANL) released 58 consecutive days of de-identified enterprise events. The data includes authentication, process, DNS, and network flow events plus labeled red team compromises.

    Known weaknesses

    Red team events are very rare compared with normal activity. De-identification removes context such as real usernames and host roles.

    Access
    Free registration
    License
    CC0
    Data type
    CSV event logs
    Size
    Not stated
    Time span
    58 days
    Link checked
    2026-09-27

    Dataset page for LANL Comprehensive, Multi-Source Cyber-Security Events (2015)Paper to cite for LANL Comprehensive, Multi-Source Cyber-Security Events (2015)

  • LANL Unified Host and Network Dataset (2017)

    Network traffic and intrusion detectionEnterprise network and authentication

    The LANL Unified Host and Network Dataset combines Windows host event logs and router flow records from one enterprise network. Host and network identities match, which supports joint host and network analysis.

    Known weaknesses

    The release contains no labeled attacks, so it suits anomaly detection and baseline modeling. Values are de-identified to protect the security of the LANL environment.

    Access
    Free registration
    License
    CC0
    Data type
    JSON host logs, CSV netflow
    Size
    Not stated
    Time span
    About 90 days
    Link checked
    2026-09-27

    Dataset page for LANL Unified Host and Network Dataset (2017)Paper to cite for LANL Unified Host and Network Dataset (2017)

  • LITNET-2020

    Network traffic and intrusion detectionNetwork intrusion detection

    LITNET-2020 is an annotated flow dataset collected from real traffic on the Lithuanian academic network. It includes normal flows and labeled attack flows for evaluating network intrusion detection methods.

    Known weaknesses

    Labels on the real traffic depend on the annotation rules the authors applied. Attack classes are much smaller than the volume of normal traffic.

    Access
    Open download
    License
    Not stated
    Data type
    NetFlow CSV
    Size
    Not stated
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for LITNET-2020Paper to cite for LITNET-2020

  • MAWI Working Group Traffic Archive

    Network traffic and intrusion detectionInternet measurement

    The MAWI (Measurement and Analysis on the WIDE Internet) archive provides daily traces from the WIDE upstream transit link. It also holds multi-day traces from yearly measurement events for internet traffic research.

    Known weaknesses

    Traces are anonymized and packet payloads are removed before release. The traces carry no attack labels, so MAWILab labels are often used alongside them.

    Access
    Open download
    License
    Not stated
    Data type
    PCAP (headers only)
    Size
    Not stated
    Time span
    1999-present
    Link checked
    2026-09-27

    Dataset page for MAWI Working Group Traffic Archive

  • MAWILab

    Network traffic and intrusion detectionInternet measurement

    MAWILab labels traffic anomalies in the MAWI archive by combining the outputs of several anomaly detectors. Researchers use it to benchmark anomaly detectors on real backbone traffic.

    Known weaknesses

    Labels come from agreement among detectors rather than verified ground truth. Anomalies that every detector misses remain unlabeled in the archive.

    Access
    Open download
    License
    Not stated
    Data type
    Anomaly labels (CSV/XML) for MAWI PCAP
    Size
    Not stated
    Time span
    2007-2024
    Link checked
    2026-09-27

    Dataset page for MAWILabPaper to cite for MAWILab

  • NSL-KDD

    Network traffic and intrusion detectionNetwork intrusion detection (legacy)

    NSL-KDD is a filtered version of the KDD Cup 1999 data that removes duplicate records. It provides connection-level features and attack labels and remains common in teaching and legacy benchmarks.

    Known weaknesses

    This dataset is outdated because its traffic comes from the 1998 DARPA evaluation. Results on NSL-KDD rarely transfer to modern network traffic or current attack types.

    Access
    Free registration
    License
    Redistribution allowed with citation
    Data type
    Feature CSV/ARFF
    Size
    Not stated
    Time span
    1998 source traffic
    Link checked
    2026-09-27

    Dataset page for NSL-KDDPaper to cite for NSL-KDD

  • OpenINTEL

    Network traffic and intrusion detectionDNS and internet measurement

    OpenINTEL performs daily active DNS measurements across large parts of the global domain name space. Researchers use it to study DNS infrastructure, abuse, and changes over time.

    Known weaknesses

    Some zones and closed datasets require a separate request or data sharing agreement. Data volume is very large and needs big data tools.

    Access
    Open download
    License
    CC BY-NC-SA 4.0 for open data
    Data type
    DNS measurement records
    Size
    Not stated
    Time span
    2015-present
    Link checked
    2026-09-27

    Dataset page for OpenINTELPaper to cite for OpenINTEL

  • SANS Internet Storm Center DShield

    Network traffic and intrusion detectionHoneypot data

    DShield aggregates firewall logs and honeypot reports submitted by volunteers worldwide. Its public API returns attack sources, targeted ports, and web honeypot data for threat analysis.

    Known weaknesses

    Sensor coverage depends on volunteers, so geographic and network coverage is uneven. The API asks users to include contact information in requests.

    Access
    API
    License
    CC BY-NC-SA 4.0
    Data type
    API (XML, JSON)
    Size
    Not stated
    Time span
    Ongoing
    Link checked
    2026-09-27

    Dataset page for SANS Internet Storm Center DShield

  • Stratosphere Malware Capture Facility Project (MCFP)

    Network traffic and intrusion detectionBotnet and malware traffic

    The Malware Capture Facility Project publishes long-term captures of real malware and normal traffic from CTU University. Each capture folder includes packet captures, flow files, and malware details for detection research.

    Known weaknesses

    Captures vary widely in length, labeling depth, and documentation quality across the archive. The directory listing gives little guidance, so users must read each capture README.

    Access
    Open download
    License
    Not stated
    Data type
    PCAP, Argus/Zeek flows
    Size
    Not stated
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for Stratosphere Malware Capture Facility Project (MCFP)

  • UCSD Network Telescope (CAIDA)

    Network traffic and intrusion detectionDarknet telescope

    The UCSD Network Telescope captures unsolicited traffic sent to a large block of unused IPv4 addresses. The traces show scanning, backscatter from spoofed DDoS attacks, and malware spread.

    Known weaknesses

    Access requires a request form and an approved research purpose. The daily data volume demands major storage and processing capacity from users.

    Access
    Academic request
    License
    CAIDA Acceptable Use Agreement and Telescope Supplement
    Data type
    PCAP, aggregated flow data
    Size
    More than 3 TB uncompressed per day
    Time span
    Ongoing, recent 14 days by request
    Link checked
    2026-09-27

    Dataset page for UCSD Network Telescope (CAIDA)

  • UGR'16

    Network traffic and intrusion detectionNetwork intrusion detection

    UGR'16 contains NetFlow v9 records from collectors inside a Spanish internet service provider (ISP). A four-month calibration set holds real background traffic, and a later test set adds synthetic attacks.

    Known weaknesses

    Later studies report labeling inaccuracies that make measured model performance uncertain. The real background traffic may also contain attacks that carry no label.

    Access
    Open download
    License
    Not stated
    Data type
    NetFlow CSV
    Size
    Not stated
    Time span
    March to August 2016
    Link checked
    2026-09-27

    Dataset page for UGR'16Paper to cite for UGR'16

  • UNSW-NB15

    Network traffic and intrusion detectionNetwork intrusion detection

    UNSW-NB15 contains 100 GB of raw traffic from the UNSW Canberra Cyber Range with nine attack families. It offers PCAP, Argus, Bro logs, and labeled CSV records for training and testing IDS models.

    Known weaknesses

    Attack traffic was generated synthetically with the IXIA PerfectStorm tool. Several attack families have very few records, which creates strong class imbalance.

    Access
    Open download
    License
    Free for academic research, commercial use requires author agreement
    Data type
    PCAP, Argus, Bro logs, CSV
    Size
    2,540,044 records, 100 GB PCAP
    Time span
    2015
    Link checked
    2026-09-27

    Dataset page for UNSW-NB15Paper to cite for UNSW-NB15

  • UQ NetFlow NIDS datasets (NF-UQ-NIDS)

    Network traffic and intrusion detectionNetwork intrusion detection

    The University of Queensland converted UNSW-NB15, BoT-IoT, ToN-IoT, and CSE-CIC-IDS2018 into a common NetFlow feature set. Versions 1 to 3 support cross-dataset evaluation of network intrusion detection systems (NIDS).

    Known weaknesses

    These are derived datasets, so they inherit synthetic traffic and labeling errors from the source datasets. The feature sets differ across versions 1, 2, and 3, so results are not directly comparable.

    Access
    Open download
    License
    Not stated
    Data type
    NetFlow CSV
    Size
    Not stated
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for UQ NetFlow NIDS datasets (NF-UQ-NIDS)Paper to cite for UQ NetFlow NIDS datasets (NF-UQ-NIDS)

  • Enron Email Dataset

    Phishing, spam, and malicious URLsEmail corpus

    Carnegie Mellon University (CMU) hosts this corpus of about 500,000 emails from about 150 Enron employees. The Federal Energy Regulatory Commission released the data during its investigation.

    Known weaknesses

    The corpus has no phishing labels, so it usually serves as legitimate email. It contains real personal information, so handle it with care.

    Access
    Open download
    License
    Not stated
    Data type
    Raw email messages in folder structure
    Size
    About 0.5 million messages, 1.7 GB compressed
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for Enron Email DatasetPaper to cite for Enron Email Dataset

  • ISCX-URL2016

    Phishing, spam, and malicious URLsMalicious URL classification

    The Canadian Institute for Cybersecurity (CIC) built this dataset of benign and malicious Uniform Resource Locators (URLs). It covers benign, spam, phishing, malware, and defacement URL classes.

    Known weaknesses

    Downloads pass through a short Canadian Institute for Cybersecurity registration form. The URL sources are older and many links are no longer active.

    Access
    Free registration
    License
    Citation required per CIC dataset terms
    Data type
    URL lists and lexical features (CSV)
    Size
    About 114,000 URLs across five classes
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for ISCX-URL2016Paper to cite for ISCX-URL2016

  • Nazario Phishing Corpus

    Phishing, spam, and malicious URLsPhishing email

    Jose Nazario maintains this corpus of hand classified phishing emails from his personal inbox. Files are grouped into mailbox archives by year, with recent files released annually.

    Known weaknesses

    The corpus is representative rather than exhaustive because it comes from one inbox. Later mailboxes are not anonymized and may contain personal data.

    Access
    Open download
    License
    CC BY 4.0
    Data type
    Email messages in mbox format
    Size
    Mailbox files from about 3 MB to 31 MB
    Time span
    2005 to 2025
    Link checked
    2026-09-27

    Dataset page for Nazario Phishing Corpus

  • OpenPhish Community Feed

    Phishing, spam, and malicious URLsPhishing URL feed

    OpenPhish publishes a free community feed of phishing Uniform Resource Locators (URLs) found by automated analysis. The community tier updates every 12 hours as a plain text list.

    Known weaknesses

    The free tier omits brand, network, and historical metadata found in paid tiers. The feed contains live malicious links that can infect systems or steal credentials. Open these links only inside an isolated sandbox or virtual machine and never on production systems.

    Access
    Open download
    License
    OpenPhish terms of use
    Data type
    Plain text list of phishing URLs
    Size
    Not stated
    Time span
    Rolling current feed
    Link checked
    2026-09-27

    Dataset page for OpenPhish Community Feed

  • Phishing Email Curated Datasets

    Phishing, spam, and malicious URLsPhishing and spam email

    This Zenodo record bundles 11 cleaned email corpora for phishing detection in consistent CSV files. It includes Enron, Ling, CEAS 2008, Nazario, Nigerian fraud, SpamAssassin, and TREC 2005 to 2007.

    Known weaknesses

    The original TREC spam corpus download links now return errors. This Zenodo record offers a practical secondary copy of those corpora for research.

    Access
    Open download
    License
    CC BY 4.0
    Data type
    Email text with sender, subject, body, and labels (CSV)
    Size
    11 CSV files, about 450 MB
    Time span
    1995 to 2022
    Link checked
    2026-09-27

    Dataset page for Phishing Email Curated DatasetsPaper to cite for Phishing Email Curated Datasets

  • Phishing Websites Dataset (Vrbancic)

    Phishing, spam, and malicious URLsPhishing website features

    This Mendeley Data set provides 88,647 website instances labeled as legitimate or phishing. Each instance carries 111 features derived from the Uniform Resource Locator (URL) and domain.

    Known weaknesses

    The full and small variants have different class balance between legitimate and phishing. Report which variant you use so others can reproduce your results.

    Access
    Open download
    License
    CC BY 4.0
    Data type
    Tabular features with binary labels (CSV)
    Size
    88,647 instances (full), 58,645 instances (small)
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for Phishing Websites Dataset (Vrbancic)Paper to cite for Phishing Websites Dataset (Vrbancic)

  • Phishpedia Phishing Benchmark

    Phishing, spam, and malicious URLsVisual phishing detection

    The Phishpedia project released a benchmark of about 30,000 phishing webpages for visual brand identification research. Each sample includes the URL, HTML, screenshot, and target brand.

    Known weaknesses

    The data is hosted on Google Drive through a link in the repository. Screenshots show real brand logos, so avoid redistributing them outside research use.

    Access
    Open download
    License
    CC0 1.0 (repository)
    Data type
    URLs, HTML, screenshots, and brand labels
    Size
    About 30,000 phishing webpages
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for Phishpedia Phishing Benchmark

  • PhishTank

    Phishing, spam, and malicious URLsPhishing URL feed

    PhishTank is a community database of phishing Uniform Resource Locators (URLs) verified by volunteers. Cisco operates the service and publishes the full database as hourly downloadable files.

    Known weaknesses

    Unregistered users are limited to a few downloads per day. The feed contains live malicious links that can infect systems or steal credentials. Open these links only inside an isolated sandbox or virtual machine and never on production systems.

    Access
    Free registration
    License
    PhishTank terms of use (Cisco)
    Data type
    URLs with submission time, verification status, and target brand (CSV, JSON, XML)
    Size
    Not stated
    Time span
    2006 to present
    Link checked
    2026-09-27

    Dataset page for PhishTank

  • PhiUSIIL Phishing URL Dataset

    Phishing, spam, and malicious URLsPhishing URL features

    PhiUSIIL contains 235,795 labeled Uniform Resource Locators (URLs) with features from page source and URL text. The collection includes 134,850 legitimate URLs and 100,945 phishing URLs with 54 features each.

    Known weaknesses

    Several features are derived from page content and may leak label information. Check each feature definition before modeling to avoid inflated accuracy.

    Access
    Open download
    License
    CC BY 4.0
    Data type
    Tabular features (54 columns) with binary labels
    Size
    235,795 URLs
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for PhiUSIIL Phishing URL DatasetPaper to cite for PhiUSIIL Phishing URL Dataset

  • PhreshPhish

    Phishing, spam, and malicious URLsPhishing webpages with HTML

    PhreshPhish pairs phishing and benign Uniform Resource Locators (URLs) with full webpage HyperText Markup Language (HTML). The release holds 666,315 samples with train and test splits.

    Known weaknesses

    Stored HTML may contain credential harvesting forms, tracking code, and malicious scripts. Do not render these pages in a browser outside an isolated sandbox.

    Access
    Open download
    License
    CC BY 4.0, anti-phishing research only
    Data type
    URL, HTML, collection date, language, and label
    Size
    666,315 samples
    Time span
    2024 to 2025
    Link checked
    2026-09-27

    Dataset page for PhreshPhishPaper to cite for PhreshPhish

  • SMS Spam Collection

    Phishing, spam, and malicious URLsSMS spam and smishing

    This University of California, Irvine (UCI) dataset holds 5,574 Short Message Service (SMS) messages labeled ham or spam. Messages come mainly from sources in the United Kingdom and in Singapore.

    Known weaknesses

    The class balance is skewed toward ham, with 747 spam messages overall. The messages predate current smishing campaigns that use links and delivery lures.

    Access
    Open download
    License
    CC BY 4.0
    Data type
    Text messages with labels
    Size
    5,574 messages
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for SMS Spam CollectionPaper to cite for SMS Spam Collection

  • SpamAssassin Public Corpus

    Phishing, spam, and malicious URLsSpam email

    The Apache SpamAssassin project published this corpus of spam, easy ham, and hard ham email. It remains a common baseline for spam filtering and phishing detection studies.

    Known weaknesses

    The messages are more than twenty years old and reflect early spam tactics. Modern spam and phishing campaigns look quite different from these samples.

    Access
    Open download
    License
    Not stated
    Data type
    Email messages in tar.bz2 archives
    Size
    Nine archives of about 1 to 2 MB each
    Time span
    2002 to 2005
    Link checked
    2026-09-27

    Dataset page for SpamAssassin Public Corpus

  • UCI Phishing Websites

    Phishing, spam, and malicious URLsPhishing website features

    This University of California, Irvine (UCI) dataset describes 11,055 websites with 30 engineered phishing indicators. Samples were gathered from PhishTank, MillerSmiles, and Google search results by the dataset authors.

    Known weaknesses

    The data predates modern phishing kits and the wide adoption of HTTPS encryption. Results from this dataset may not transfer to current phishing attacks.

    Access
    Open download
    License
    CC BY 4.0
    Data type
    Tabular integer features with binary labels
    Size
    11,055 instances, 30 features
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for UCI Phishing WebsitesPaper to cite for UCI Phishing Websites

  • URLhaus

    Phishing, spam, and malicious URLsMalware distribution URLs

    URLhaus from abuse.ch tracks Uniform Resource Locators (URLs) used to distribute malware. It offers database dumps every five minutes plus a free community Application Programming Interface (API).

    Known weaknesses

    All downloads require a free Auth-Key from the abuse.ch authentication portal. The feed contains live malicious links that can infect systems or steal credentials. Open these links only inside an isolated sandbox or virtual machine and never on production systems.

    Access
    Free registration
    License
    abuse.ch fair use terms, commercial heavy use may require a paid subscription
    Data type
    Malicious URLs with tags, status, and payload hashes (CSV, JSON, MISP, RPZ)
    Size
    Not stated
    Time span
    2018 to present
    Link checked
    2026-09-27

    Dataset page for URLhaus

  • Web Page Phishing Detection (Hannousse and Yahiouche)

    Phishing, spam, and malicious URLsPhishing website features

    This Mendeley Data set holds 11,430 balanced phishing and legitimate Uniform Resource Locators (URLs). Each URL has 87 features drawn from URL syntax, page content, and external services.

    Known weaknesses

    External service features such as search rank may not be reproducible today. Recompute these features from current sources before reusing the dataset.

    Access
    Open download
    License
    CC BY 4.0
    Data type
    URLs, Document Object Model trees, and tabular features
    Size
    11,430 URLs, 87 features
    Time span
    Collected May 2020
    Link checked
    2026-09-27

    Dataset page for Web Page Phishing Detection (Hannousse and Yahiouche)Paper to cite for Web Page Phishing Detection (Hannousse and Yahiouche)

  • AnnoCTR

    Security knowledge bases and textEntity linking in threat reports

    AnnoCTR contains cyber threat reports annotated with named entities, temporal expressions, and security concepts. Entities and concepts link to Wikipedia and to MITRE ATT&CK tactics and techniques.

    Known weaknesses

    The repository is an archived research prototype that the authors do not maintain. Share-alike license terms apply to any derived datasets that researchers distribute.

    Access
    Open download
    License
    CC-BY-SA-4.0
    Data type
    Annotated text files
    Size
    400 cyber threat reports
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for AnnoCTRPaper to cite for AnnoCTR

  • APTNER

    Security knowledge bases and textNamed entity recognition

    APTNER is a named entity recognition dataset built from cyber threat intelligence reports and security blogs. It defines 21 entity types that align with common threat intelligence concepts.

    Known weaknesses

    The repository has no license and has not been maintained since 2022. Its label definitions differ from DNRTI and CyNER, so merging datasets needs a label mapping.

    Access
    Open download
    License
    Not stated
    Data type
    Token-labeled text files
    Size
    Train, development, and test files of about 2.4 MB
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for APTNERPaper to cite for APTNER

  • CASIE

    Security knowledge bases and textCybersecurity event extraction

    CASIE (CyberAttack Sensing and Information Extraction) provides news articles annotated with cybersecurity events and arguments. Event types include data breach, phishing, ransom, vulnerability discovery, and patch.

    Known weaknesses

    The repository has no license file, so reuse terms should be confirmed with the authors. News articles differ in style and detail from technical threat intelligence reports.

    Access
    Open download
    License
    Not stated
    Data type
    JSON annotation and source text files
    Size
    1,000 annotated articles
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for CASIEPaper to cite for CASIE

  • CTIBench

    Security knowledge bases and textLLM benchmark for threat intelligence

    CTIBench evaluates LLMs on practical cyber threat intelligence tasks drawn from real analyst work. Tasks cover CTI knowledge, CVE to CWE mapping, CVSS prediction, ATT&CK extraction, and actor attribution.

    Known weaknesses

    Some tasks use public CVE and report text that may already appear in model training data. The repository moved from its original GitHub owner, so older links may redirect.

    Access
    Open download
    License
    Not stated
    Data type
    TSV task files
    Size
    4,610 benchmark examples across five task families
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for CTIBenchPaper to cite for CTIBench

  • CTID Mappings Explorer

    Security knowledge bases and textControl and vulnerability to ATT&CK mappings

    The Center for Threat-Informed Defense (CTID) maps security controls and capabilities to ATT&CK techniques. The project also absorbed the earlier CTID mappings from CVE records to ATT&CK.

    Known weaknesses

    Mappings reflect analyst judgment and cover only selected ATT&CK versions. Vulnerability mappings cover a small subset of CVEs rather than the full CVE List.

    Access
    Open download
    License
    Apache-2.0
    Data type
    JSON, CSV, and Excel mapping files
    Size
    Not stated
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for CTID Mappings Explorer

  • CyberMetric

    Security knowledge bases and textLLM benchmark for security knowledge

    CyberMetric is a set of multiple-choice questions on cybersecurity knowledge for evaluating language models. Questions were generated with retrieval from standards, papers, and books, then validated by experts.

    Known weaknesses

    Questions were generated with GPT-3.5, so wording may favor models from similar families. The authors report expert validation, but errors may remain in the larger subsets.

    Access
    Open download
    License
    Not stated
    Data type
    JSON question files
    Size
    Subsets of 80, 500, 2,000, and 10,000 questions
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for CyberMetricPaper to cite for CyberMetric

  • CyNER

    Security knowledge bases and textNamed entity recognition

    CyNER is a Python library for cybersecurity named entity recognition (NER) in threat reports. Its repository includes annotated sentences split into train, validation, and test files.

    Known weaknesses

    The annotated corpus is small and follows its own entity labeling scheme. Results may not transfer to CTI NER datasets that use different label sets.

    Access
    Open download
    License
    MIT
    Data type
    CoNLL-style text files
    Size
    About 1 MB of annotated text
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for CyNERPaper to cite for CyNER

  • DNRTI

    Security knowledge bases and textNamed entity recognition

    DNRTI is a named entity recognition dataset built from more than 300 threat intelligence reports. Experts annotated entities in 13 categories such as hacker organizations, tools, and attack actions.

    Known weaknesses

    The repository has no license and offers minimal documentation beyond the archive itself. Users must check the split files and label conventions before comparing results across papers.

    Access
    Open download
    License
    Not stated
    Data type
    Token-labeled text in a RAR archive
    Size
    6,574 sentences with 36,412 entities
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for DNRTIPaper to cite for DNRTI

  • MalwareTextDB

    Security knowledge bases and textMalware report annotation

    MalwareTextDB annotates APT reports with malware entities, relations, and attribute labels from Malware Attribute Enumeration and Characterization (MAEC). The corpus was later extended for the SemEval-2018 Task 8 shared task.

    Known weaknesses

    The corpus is small and its reports date from 2017 or earlier. The SemEval extension covering 85 reports is distributed separately from this archive.

    Access
    Open download
    License
    Not stated
    Data type
    Annotated text in a zip archive
    Size
    39 APT reports with 6,819 sentences (v1.0)
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for MalwareTextDBPaper to cite for MalwareTextDB

  • MITRE ATLAS Data

    Security knowledge bases and textAdversarial machine learning threats

    MITRE ATLAS (Adversarial Threat Landscape for Artificial-Intelligence Systems) describes adversary tactics and techniques against AI systems. The data also includes documented case studies of attacks on machine learning systems.

    Known weaknesses

    The field changes quickly, so technique lists shift noticeably between ATLAS versions. Case studies include research demonstrations as well as incidents observed in real deployments.

    Access
    Open download
    License
    Apache-2.0
    Data type
    YAML
    Size
    170 techniques and sub-techniques and 57 case studies (v5.6.0)
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for MITRE ATLAS Data

  • MITRE ATT&CK STIX Data

    Security knowledge bases and textAdversary tactics and techniques

    MITRE ATT&CK describes adversary tactics, techniques, groups, and software based on real-world observations. This repository provides the Enterprise, Mobile, and Industrial Control Systems (ICS) domains as Structured Threat Information Expression (STIX) 2.1 bundles.

    Known weaknesses

    Technique identifiers change between versions through revocation, deprecation, and restructuring. Studies should record the exact ATT&CK version used for every mapping or label.

    Access
    Open download
    License
    MITRE ATT&CK terms of use
    Data type
    STIX 2.1 JSON
    Size
    ATT&CK v19.2 for Enterprise, Mobile, and ICS
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for MITRE ATT&CK STIX Data

  • MITRE CAPEC

    Security knowledge bases and textAttack pattern taxonomy

    CAPEC catalogs common attack patterns that adversaries use against software, hardware, and communication systems. Entries link to related CWE weaknesses and include prerequisites and mitigations.

    Known weaknesses

    CAPEC receives updates less often than ATT&CK and can lag behind current attacker behavior. Some attack patterns are abstract and hard to map to observed security events.

    Access
    Open download
    License
    CAPEC terms of use
    Data type
    XML and CSV
    Size
    CAPEC List Version 3.9
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for MITRE CAPEC

  • MITRE CTI Repository

    Security knowledge bases and textAdversary tactics and attack patterns

    The MITRE cyber threat intelligence (CTI) repository provides ATT&CK content in the STIX 2.0 format. It also includes Common Attack Pattern Enumeration and Classification (CAPEC) data for tools built on python-stix2.

    Known weaknesses

    This repository uses STIX 2.0 and lacks the collection features of the newer repository. New projects should prefer attack-stix-data when they need STIX 2.1 ATT&CK content.

    Access
    Open download
    License
    MITRE terms of use
    Data type
    STIX 2.0 JSON
    Size
    Not stated
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for MITRE CTI Repository

  • MITRE CWE

    Security knowledge bases and textWeakness taxonomy

    CWE is a hierarchical list of common software and hardware weakness types maintained by MITRE. It supplies the weakness labels used by NVD and most vulnerable code datasets.

    Known weaknesses

    CWE labels assigned in NVD records are often coarse, generic, or inconsistent across analysts. Researchers should choose one CWE view, such as View-1003, to aggregate labels consistently.

    Access
    Open download
    License
    CWE terms of use (royalty-free)
    Data type
    XML and CSV
    Size
    CWE List Version 4.20
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for MITRE CWE

  • MITRE D3FEND

    Security knowledge bases and textDefensive technique ontology

    D3FEND is a knowledge graph of cybersecurity countermeasures and the digital artifacts they act on. It links defensive techniques to ATT&CK offensive techniques through those shared artifacts.

    Known weaknesses

    Mappings between defensive and offensive techniques are inferred through artifacts rather than observed outcomes. The ontology structure changes across releases, so pin a specific version.

    Access
    Open download
    License
    MIT
    Data type
    OWL ontology, JSON-LD, and CSV
    Size
    Not stated
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for MITRE D3FEND

  • Primus-FineWeb

    Security knowledge bases and textSecurity pretraining corpus

    Primus-FineWeb is a cybersecurity text corpus filtered from the FineWeb web crawl with a trained classifier. It supports continued pretraining of language models for security tasks.

    Known weaknesses

    Access requires a Hugging Face login and agreement to share contact information. Users must also follow the FineWeb license and the Common Crawl terms of use.

    Access
    Free registration
    License
    ODC-By
    Data type
    Text records in a Hugging Face dataset
    Size
    2.57 billion tokens
    Time span
    Up to CC-MAIN-2024-10
    Link checked
    2026-09-27

    Dataset page for Primus-FineWebPaper to cite for Primus-FineWeb

  • Sigma Rules

    Security knowledge bases and textDetection rules

    Sigma is a generic signature format for writing log-based detection rules across security platforms. The main repository holds community detection rules tagged with ATT&CK techniques.

    Known weaknesses

    Rule quality and false positive rates vary by author and by log source. The Detection Rule License (DRL) requires attribution when rules are reused or converted.

    Access
    Open download
    License
    Detection Rule License (DRL) 1.1
    Data type
    YAML rule files
    Size
    More than 3,000 detection rules
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for Sigma Rules

  • TRAM Training Data

    Security knowledge bases and textReport to ATT&CK mapping

    The Threat Report ATT&CK Mapper (TRAM) repository includes annotated CTI reports for training technique classifiers. The shipped model identifies up to 50 common ATT&CK techniques in report text.

    Known weaknesses

    The training data covers a limited technique subset and targets ATT&CK version 13. Newer ATT&CK releases have renamed or restructured some techniques used in these labels.

    Access
    Open download
    License
    Apache-2.0
    Data type
    JSON annotation files
    Size
    Not stated
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for TRAM Training Data

  • abuse.ch Feodo Tracker

    Threat intelligence and hacker communitiesBotnet command and control feed

    Feodo Tracker lists botnet command and control (C2) servers for malware families that grew from banking trojans. Blocklists include IP addresses with first seen and last online dates.

    Known weaknesses

    Blocklists show currently tracked C2 servers, so historical research needs regular snapshots. The abuse.ch fair use limits and commercial terms apply to all automated access.

    Access
    Free registration
    License
    abuse.ch terms of use
    Data type
    CSV, JSON, and text blocklists
    Size
    Not stated
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for abuse.ch Feodo Tracker

  • abuse.ch SSL Blacklist (SSLBL)

    Threat intelligence and hacker communitiesMalicious TLS certificates

    The abuse.ch blocklist SSLBL lists Secure Sockets Layer (SSL) certificates used by botnet C2 servers and malware. Entries include certificate SHA1 fingerprints together with the associated malware family names.

    Known weaknesses

    Coverage reflects what abuse.ch observes and does not include all malicious certificates. The abuse.ch fair use limits and commercial terms apply to all automated access.

    Access
    Free registration
    License
    abuse.ch terms of use
    Data type
    CSV and text blocklists
    Size
    Not stated
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for abuse.ch SSL Blacklist (SSLBL)

  • abuse.ch ThreatFox

    Threat intelligence and hacker communitiesIndicators of compromise

    ThreatFox is an abuse.ch platform for sharing indicators of compromise (IOCs) linked to malware. Indicators include IP addresses, domains, URLs, and file hashes with malware family tags.

    Known weaknesses

    Access requires an abuse.ch account and a personal Auth-Key for API queries. Indicators are submitted by the community and can include false positives or stale entries. Commercial use requires a paid subscription that Spamhaus manages on behalf of abuse.ch.

    Access
    Free registration
    License
    abuse.ch terms of use
    Data type
    CSV and JSON exports plus REST API
    Size
    Not stated
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for abuse.ch ThreatFox

  • abuse.ch URLhaus

    Threat intelligence and hacker communitiesMalware URL feed

    URLhaus is an abuse.ch project that collects URLs used to distribute malware. Entries include online status, tags, payload hashes, and the reporter who submitted each URL.

    Known weaknesses

    Access now requires an authenticated abuse.ch account and a personal Auth-Key for queries. Free use covers non-commercial volumes, and commercial use needs a paid Spamhaus subscription.

    Access
    Free registration
    License
    abuse.ch terms of use
    Data type
    CSV, JSON, and text exports plus REST API
    Size
    Not stated
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for abuse.ch URLhaus

  • APT and Cybercriminal Campaign Collections (CyberMonitor)

    Threat intelligence and hacker communitiesAdvanced persistent threat reports

    CyberMonitor collects public reports on APT and cybercriminal campaigns organized by publication year. The repository stores report files alongside an index in its main README file.

    Known weaknesses

    The collection was last updated in July 2024 and may not receive further additions. Reports remain under their publishers' copyrights, and some files may trigger antivirus alerts.

    Access
    Open download
    License
    Not stated
    Data type
    PDF reports with Markdown index
    Size
    Not stated
    Time span
    2006-2024
    Link checked
    2026-09-27

    Dataset page for APT and Cybercriminal Campaign Collections (CyberMonitor)

  • APTnotes

    Threat intelligence and hacker communitiesAdvanced persistent threat reports

    APTnotes indexes public vendor reports on campaigns attributed to advanced persistent threat (APT) groups. A CSV file lists the title, source, link, hash, and date for each report.

    Known weaknesses

    Report PDFs are hosted on Box and require the separate download tools in aptnotes/tools. Vendor reports carry their own copyrights, and the index receives only infrequent updates.

    Access
    Open download
    License
    Not stated
    Data type
    CSV and JSON index with PDF reports
    Size
    689 reports
    Time span
    2006-2024
    Link checked
    2026-09-27

    Dataset page for APTnotes

  • Cambridge Cybercrime Centre CrimeBB

    Threat intelligence and hacker communitiesUnderground forum corpus

    The Cambridge Cybercrime Centre shares CrimeBB, a collection of posts scraped from underground cybercrime forums. Related collections cover marketplace contracts, chat channels, and website defacements.

    Known weaknesses

    Institutional review board (IRB) review and an agreement signed by the researcher's institution are required. Researchers must send the Centre a copy of each paper when it is submitted for peer review. The Centre places significant restrictions on any commercial use of its data.

    Access
    Academic request
    License
    Cambridge Cybercrime Centre data sharing agreement
    Data type
    Database of forum posts, threads, and user records
    Size
    More than 100 million forum posts
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for Cambridge Cybercrime Centre CrimeBBPaper to cite for Cambridge Cybercrime Centre CrimeBB

  • CyberTweets

    Threat intelligence and hacker communitiesSocial media threat intelligence

    CyberTweets is a corpus of tweets annotated for cyber threat relevance and threat type. The repository also includes the web application used by annotators to label the tweets.

    Known weaknesses

    The tweets predate 2019 and may reference accounts or posts that no longer exist. Any reuse must follow the X platform terms for redistributing content.

    Access
    Open download
    License
    Not stated
    Data type
    MongoDB dump
    Size
    About 21,000 annotated tweets
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for CyberTweetsPaper to cite for CyberTweets

  • Evolution Cryptomarket Structured Dataset

    Threat intelligence and hacker communitiesUnderground marketplace and forum

    This dataset structures forum and marketplace data from the Evolution dark web market into relational tables. It also provides temporal communication networks built from forum interactions between users.

    Known weaknesses

    Content describes illicit drug and carding trade by pseudonymous users of the market. IRB review is advised before researchers analyze or publish results from these data. Zenodo blocks some automated clients, so researchers may need to download through a browser.

    Access
    Open download
    License
    CC-BY-4.0
    Data type
    Relational tables and network snapshots
    Size
    More than 500,000 forum posts and 80,000 listings
    Time span
    2014-01 to 2015-03
    Link checked
    2026-09-27

    Dataset page for Evolution Cryptomarket Structured DatasetPaper to cite for Evolution Cryptomarket Structured Dataset

  • LevelBlue Open Threat Exchange (OTX)

    Threat intelligence and hacker communitiesCommunity threat intelligence

    The Open Threat Exchange (OTX), formerly branded AlienVault OTX, is a free community threat intelligence platform. Users share pulses that group IOCs with context about a threat.

    Known weaknesses

    Pulses are user generated and vary widely in quality, accuracy, and completeness. Researchers should deduplicate indicators and check pulse authorship before any analysis.

    Access
    Free registration
    License
    OTX terms of service
    Data type
    JSON via DirectConnect API and SDKs
    Size
    More than 100,000 participants in 140 countries per the provider
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for LevelBlue Open Threat Exchange (OTX)

  • Malpedia

    Threat intelligence and hacker communitiesMalware family knowledge base

    Malpedia from Fraunhofer FKIE documents malware families and threat actors with curated report references. Public API endpoints list families, actors, and a bibliography of analysis reports.

    Known weaknesses

    Malware samples and some YARA rules require an account in an invite-only trust group. Family and actor naming follows Malpedia conventions that may differ from vendor names.

    Access
    API
    License
    Not stated
    Data type
    JSON via REST API and BibTeX
    Size
    3,818 malware families and 1,056 actors
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for Malpedia

  • MISP Default Feeds

    Threat intelligence and hacker communitiesOpen source threat feeds

    The Malware Information Sharing Platform (MISP) project lists public open source intelligence feeds enabled by default. The feeds are distributed in MISP, CSV, and freetext formats for automated import.

    Known weaknesses

    Each feed has its own provider terms, and at least one listed feed requires a paid license. Researchers should check each source's terms before any redistribution of feed data.

    Access
    Open download
    License
    Varies by feed provider
    Data type
    MISP JSON, CSV, and freetext feeds
    Size
    About 90 listed feeds
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for MISP Default Feeds

  • MISP Galaxy

    Threat intelligence and hacker communitiesThreat actor and malware taxonomies

    MISP Galaxy provides clusters that describe threat actors, malware families, tools, and related concepts. Clusters include synonyms and relationships that help link intelligence from different vendors.

    Known weaknesses

    Community contributors maintain the clusters, so actor attributions can conflict with vendor naming. Synonym lists need manual checking before researchers merge entities across sources.

    Access
    Open download
    License
    CC0-1.0 (dual-licensed)
    Data type
    JSON cluster files
    Size
    Not stated
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for MISP Galaxy

  • ORKL Threat Report Library

    Threat intelligence and hacker communitiesThreat report library

    ORKL is a searchable library of threat intelligence reports with extracted text and threat actor links. A public REST API exposes library entries, threat actor records, and library statistics.

    Known weaknesses

    The web interface is a client-rendered application, so bulk access should use the documented API. Report content remains under the rights of the original publishing vendors.

    Access
    API
    License
    Not stated
    Data type
    JSON via REST API
    Size
    29,534 library entries and 1,945 threat actor entries
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for ORKL Threat Report Library

  • PEASEC Cybersecurity Dataset

    Threat intelligence and hacker communitiesSecurity text corpus

    This corpus combines security tweets, blog posts, arXiv security papers, and CVE descriptions. The authors built it to train the CySecBERT domain-adapted language model.

    Known weaknesses

    The repository shares references only because of data storage and publication restrictions. Researchers must rehydrate tweets and blog text, and some items are no longer available.

    Access
    Open download
    License
    Not stated
    Data type
    Identifiers and collection scripts
    Size
    4.3 million entries and 528 million tokens
    Time span
    Up to 2022-03
    Link checked
    2026-09-27

    Dataset page for PEASEC Cybersecurity DatasetPaper to cite for PEASEC Cybersecurity Dataset

  • Ransomware.live

    Threat intelligence and hacker communitiesRansomware leak site tracking

    Ransomware.live tracks victim claims that ransomware groups post on their leak sites. It reports victim metadata, group profiles, and statistics through a website and a public API.

    Known weaknesses

    Victim claims come from criminal leak sites and are not independently verified. Records name real organizations, so researchers should handle and publish them with care. A separate paid Pro API offers additional features beyond the free public API.

    Access
    API
    License
    Not stated
    Data type
    JSON via REST API
    Size
    32,124 victims and 402 groups tracked
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for Ransomware.live

  • Spamhaus Don't Route Or Peer (DROP)

    Threat intelligence and hacker communitiesMalicious network blocklist

    Spamhaus Don't Route Or Peer (DROP) lists IP ranges that Spamhaus considers too dangerous to route. Separate files cover IPv4 ranges, IPv6 ranges, and autonomous system numbers (ASNs).

    Known weaknesses

    Spamhaus asks users not to fetch the list automatically more than once per hour. The Extended DROP (EDROP) list merged into DROP on April 10, 2024. Products that use DROP must credit Spamhaus and keep the file copyright text.

    Access
    Open download
    License
    Free use with Spamhaus attribution
    Data type
    JSON files
    Size
    Not stated
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for Spamhaus Don't Route Or Peer (DROP)

  • CISA Known Exploited Vulnerabilities (KEV) Catalog

    Vulnerabilities and exploitsExploited vulnerabilities

    The Cybersecurity and Infrastructure Security Agency (CISA) lists CVEs with confirmed exploitation in the wild. Each entry includes vendor, product, date added, required action, and a known ransomware use flag.

    Known weaknesses

    The catalog includes only exploitation that CISA has confirmed through its own evidence standards. Absence from KEV does not mean that a vulnerability has never been exploited. A mirror with full commit history is available on GitHub at cisagov/kev-data.

    Access
    Open download
    License
    CC0-1.0
    Data type
    CSV and JSON with schema
    Size
    1,726 entries (catalog version 2026.09.25)
    Time span
    2021-11-03 to present
    Link checked
    2026-09-27

    Dataset page for CISA Known Exploited Vulnerabilities (KEV) Catalog

  • CISA Vulnrichment

    Vulnerabilities and exploitsVulnerability enrichment

    CISA acts as an Authorized Data Publisher (ADP) that adds enrichment to selected CVE records. Added fields include Stakeholder-Specific Vulnerability Categorization (SSVC) decision points, CWE, CVSS, and CPE data.

    Known weaknesses

    CISA enriches only a subset of CVEs, so coverage is uneven across vendors and years. The same ADP data also appears inside official CVE Records in the cvelistV5 repository.

    Access
    Open download
    License
    CC0-1.0
    Data type
    JSON files in a Git repository
    Size
    Not stated
    Time span
    2024-present
    Link checked
    2026-09-27

    Dataset page for CISA Vulnrichment

  • CVE List V5 (cvelistV5)

    Vulnerabilities and exploitsVulnerability database

    The CVE Program repository holds every CVE Record in the CVE JSON 5 record format. Files update about every seven minutes from the official CVE Services API.

    Known weaknesses

    Records mix schema versions 5.0, 5.1, and later, so validate each record against its dataversion field. Legacy CSV, HTML, and XML downloads ended on June 30, 2024 and are no longer updated.

    Access
    Open download
    License
    CVE Program Terms of Use
    Data type
    JSON files in a Git repository
    Size
    All published CVE Records
    Time span
    1999-present
    Link checked
    2026-09-27

    Dataset page for CVE List V5 (cvelistV5)

  • ENISA European Union Vulnerability Database (EUVD)

    Vulnerabilities and exploitsVulnerability database

    The European Union Agency for Cybersecurity (ENISA) runs the European Union Vulnerability Database (EUVD) under the NIS2 Directive. It aggregates records from computer security incident response teams (CSIRTs), vendors, and the CVE Program.

    Known weaknesses

    The service launched in 2025, so its coverage and data fields are still changing. EUVD identifiers differ from CVE identifiers and need a mapping step before joins.

    Access
    API
    License
    Not stated
    Data type
    JSON via REST API and web search
    Size
    Not stated
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for ENISA European Union Vulnerability Database (EUVD)

  • Exploit Database (Exploit-DB)

    Vulnerabilities and exploitsPublic exploits

    Exploit-DB is an archive of public exploits and proof-of-concept code maintained by the company OffSec. The GitLab repository includes the exploit files and a CSV index that lists CVE codes.

    Known weaknesses

    The archive contains working exploit code, so researchers should handle it inside an isolated environment. Antivirus tools may quarantine some files after download and corrupt a local copy. CVE mappings are missing for many older entries and need manual linking.

    Access
    Open download
    License
    GPL-2.0
    Data type
    Source code files and CSV index
    Size
    47,160 exploit entries
    Time span
    1988-2026
    Link checked
    2026-09-27

    Dataset page for Exploit Database (Exploit-DB)

  • FIRST Exploit Prediction Scoring System (EPSS)

    Vulnerabilities and exploitsExploit prediction scores

    The Exploit Prediction Scoring System (EPSS) estimates the probability that a CVE is exploited in the next 30 days. Scores and percentiles are published daily for every published CVE through CSV files and an API.

    Known weaknesses

    Scores produced by different EPSS model versions are not directly comparable over time. EPSS predicts the likelihood of exploitation activity and does not measure impact or severity. Studies of score changes need the archived daily files rather than the current snapshot.

    Access
    Open download
    License
    Not stated
    Data type
    Daily CSV files and JSON API
    Size
    380,066 CVEs scored on 2026-09-27
    Time span
    2021-present
    Link checked
    2026-09-27

    Dataset page for FIRST Exploit Prediction Scoring System (EPSS)Paper to cite for FIRST Exploit Prediction Scoring System (EPSS)

  • GitHub Advisory Database

    Vulnerabilities and exploitsOpen source package advisories

    GitHub publishes security advisories for open source packages in OSV format under a Creative Commons license. The database separates GitHub-reviewed advisories from unreviewed advisories imported from NVD.

    Known weaknesses

    Only GitHub-reviewed advisories include curated package names and affected version ranges. Unreviewed advisories are mostly NVD imports that lack package ecosystem data.

    Access
    Open download
    License
    CC-BY-4.0
    Data type
    JSON files in OSV format
    Size
    About 36,000 reviewed and 342,000 unreviewed advisories
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for GitHub Advisory Database

  • Metasploit Framework module metadata

    Vulnerabilities and exploitsExploit modules

    Rapid7 maintains Metasploit modules and a machine-readable metadata file inside the framework repository. The metadata lists module type, CVE references, targets, platforms, and disclosure dates.

    Known weaknesses

    Module coverage reflects Rapid7 and community priorities rather than all exploited vulnerabilities. Researchers should analyze the file db/modules_metadata_base.json instead of running any modules.

    Access
    Open download
    License
    BSD-3-Clause
    Data type
    JSON metadata and Ruby source code
    Size
    7,181 modules including 2,698 exploit modules
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for Metasploit Framework module metadata

  • NVD CVE API 2.0

    Vulnerabilities and exploitsVulnerability database

    The National Vulnerability Database (NVD) API returns Common Vulnerabilities and Exposures (CVE) records with enrichment data. Enrichment adds Common Vulnerability Scoring System (CVSS) scores, Common Weakness Enumeration (CWE) types, and Common Platform Enumeration (CPE) names. The National Institute of Standards and Technology (NIST) also publishes bulk JSON feeds for offline analysis.

    Known weaknesses

    NVD has carried a growing enrichment backlog since early 2024 because CVE submissions rose sharply. Since April 15, 2026, NIST prioritizes Known Exploited Vulnerabilities (KEV), federal, and critical-software CVEs. Unenriched CVEs published before March 1, 2026 are now marked as Not Scheduled for enrichment. A free API key raises the request rate limit compared with anonymous access.

    Access
    API
    License
    NVD terms of use
    Data type
    JSON records via REST API and bulk feeds
    Size
    All published CVE records
    Time span
    1999-present
    Link checked
    2026-09-27

    Dataset page for NVD CVE API 2.0

  • OpenSSF Malicious Packages

    Vulnerabilities and exploitsSoftware supply chain

    The Open Source Security Foundation (OpenSSF) collects reports of malicious packages found in open source registries. Reports use the OSV format and cover ecosystems such as npm and PyPI.

    Known weaknesses

    Records describe real malicious packages, so never install listed versions outside a sandbox. Coverage depends on contributing scanners and community reports rather than full registry audits.

    Access
    Open download
    License
    Apache-2.0
    Data type
    JSON reports in OSV format
    Size
    Not stated
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for OpenSSF Malicious Packages

  • OSV.dev Open Source Vulnerabilities

    Vulnerabilities and exploitsOpen source package advisories

    The OSV.dev service aggregates open source vulnerability records in the shared Open Source Vulnerability (OSV) schema. It covers more than 40 package and operating system ecosystems with affected version ranges.

    Known weaknesses

    Each upstream source carries its own license, so attribution rules must be checked per ecosystem. Some records are converted from NVD and inherit its enrichment gaps and delays.

    Access
    Open download
    License
    Varies by source (CC-BY-4.0, CC0-1.0, MIT, Apache-2.0, CC-BY-SA-4.0)
    Data type
    JSON records via Google Cloud Storage zip files and REST API
    Size
    More than 40 ecosystems from 19 direct sources
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for OSV.dev Open Source Vulnerabilities

  • VulZoo

    Vulnerabilities and exploitsVulnerability intelligence aggregation

    VulZoo integrates 17 vulnerability information sources into one dataset for security research. It links CVEs with exploits, advisories, patches, and threat knowledge bases for downstream tasks.

    Known weaknesses

    The repository has not received updates since March 2025, so bundled data is dated. Users must run the provided scripts to refresh upstream sources before new analyses.

    Access
    Open download
    License
    Not stated
    Data type
    Raw and processed files in a Git repository
    Size
    17 sources, 25 GB or more when fully cloned
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for VulZooPaper to cite for VulZoo

  • ARVO

    Vulnerable code and software securityReproducible vulnerabilities

    ARVO (Atlas of Reproducible Vulnerabilities) packages OSS-Fuzz vulnerabilities with Docker images, proof-of-concept inputs, and fix commits. Each record includes commands to rebuild the vulnerable and fixed versions of the code.

    Known weaknesses

    Docker images need large disk space and a working container runtime. Coverage is limited to open source projects that Google fuzzes through OSS-Fuzz.

    Access
    Open download
    License
    Not stated
    Data type
    SQLite metadata database and Docker images
    Size
    More than 6,100 vulnerabilities across 311 projects
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for ARVOPaper to cite for ARVO

  • Big-Vul

    Vulnerable code and software securityFunction-level vulnerability detection

    Big-Vul links CVE entries to C and C++ fix commits with code before and after each change. Each row has 21 features including CWE, CVSS subscores, and the developer commit message.

    Known weaknesses

    Commit-based labels treat every changed function as vulnerable, which adds substantial label noise. The PrimeVul paper documents low label accuracy and duplication in this dataset.

    Access
    Open download
    License
    MIT
    Data type
    CSV of functions and commits
    Size
    3,754 vulnerabilities from 348 projects across 91 CWE types
    Time span
    2002-2019
    Link checked
    2026-09-27

    Dataset page for Big-VulPaper to cite for Big-Vul

  • CleanVul

    Vulnerable code and software securityFunction-level vulnerability detection

    CleanVul uses an LLM scorer with heuristic filters to keep genuine vulnerability-fixing function changes. Each row pairs a function before and after the fix with a confidence score.

    Known weaknesses

    Labels depend on LLM judgments, and the reported correctness rates are estimates from the authors. The repository has no stated license, so reuse terms should be confirmed with the authors.

    Access
    Open download
    License
    Not stated
    Data type
    CSV function pairs with confidence scores
    Size
    8,092 pairs at 90.6% correctness (threshold 3) or 6,051 at 97.3% (threshold 4)
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for CleanVulPaper to cite for CleanVul

  • CrossVul

    Vulnerable code and software securityMulti-language vulnerability fix files

    CrossVul pairs vulnerable source files with their patched versions across many programming languages. A supporting dataset adds commit messages from the Git commits that served as security patches.

    Known weaknesses

    Labels are assigned at file level, so the specific vulnerable lines are not marked. Zenodo blocks some automated clients, so researchers may need to download through a browser.

    Access
    Open download
    License
    CC-BY-4.0
    Data type
    Source files and commit message data in a zip archive
    Size
    Files in more than 40 languages, about 1.4 GB
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for CrossVulPaper to cite for CrossVul

  • CVEfixes

    Vulnerable code and software securityVulnerability fix commits

    CVEfixes automatically collects CVE records and their fix commits from open source repositories. It stores code before and after each change at file and method level in a relational database.

    Known weaknesses

    Fix commits often include changes unrelated to the vulnerability, which creates noisy function labels. The database file is large and is distributed through Zenodo rather than GitHub.

    Access
    Open download
    License
    Data CC-BY-4.0, code MIT
    Data type
    SQLite relational database
    Size
    11,873 CVEs, 12,107 fix commits, 4,249 projects, 272 CWE types (v1.0.8)
    Time span
    Up to 2024-07-23
    Link checked
    2026-09-27

    Dataset page for CVEfixesPaper to cite for CVEfixes

  • CyberGym

    Vulnerable code and software securityAI agent vulnerability benchmark

    CyberGym tests artificial intelligence (AI) agents on generating proof-of-concept inputs for real vulnerabilities. Tasks come from open source projects with pre-patch and post-patch builds.

    Known weaknesses

    The maintainers warn users to deploy everything locally and never expose it to the internet. Full server data is very large, so storage needs careful planning in advance.

    Access
    Open download
    License
    Apache-2.0
    Data type
    Benchmark tasks, binaries, and Docker environments
    Size
    1,507 vulnerabilities across 188 projects, about 240 GB
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for CyberGymPaper to cite for CyberGym

  • CyberSecEval (Purple Llama)

    Vulnerable code and software securitySecure code generation benchmark

    CyberSecEval from Meta measures whether large language models (LLMs) generate insecure code as coding assistants. It includes autocomplete and instruction tests scored by an insecure code detector.

    Known weaknesses

    Only the insecure code tests fit software security research on vulnerable code. Other suites in the repository measure attack compliance and prompt injection instead. Rule-based detection can flag safe code as insecure and inflate reported risk.

    Access
    Open download
    License
    Not stated
    Data type
    JSON test prompts and evaluation code
    Size
    Not stated
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for CyberSecEval (Purple Llama)Paper to cite for CyberSecEval (Purple Llama)

  • Devign

    Vulnerable code and software securityFunction-level vulnerability detection

    Devign provides manually labeled vulnerability-fixing commits and C functions from large open source projects. The public release covers FFmpeg and QEMU with binary labels for each function.

    Known weaknesses

    Only two of the four studied projects are publicly released with function labels. Later studies report duplication and unrealistic class balance that inflate model scores.

    Access
    Open download
    License
    Not stated
    Data type
    JSON functions and CSV commit labels
    Size
    FFmpeg and QEMU subsets of a four-project study
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for DevignPaper to cite for Devign

  • DiverseVul

    Vulnerable code and software securityFunction-level vulnerability detection

    DiverseVul contains vulnerable and non-vulnerable C and C++ functions extracted from security fix commits. It added 295 projects beyond all prior datasets combined at its release.

    Known weaknesses

    The authors document label noise from commit-based labeling in their own manual analysis. Files are hosted on Google Drive rather than in a versioned research archive.

    Access
    Open download
    License
    Not stated
    Data type
    JSON functions with commit metadata
    Size
    18,945 vulnerable and 330,492 non-vulnerable functions from 7,514 commits across 150 CWEs
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for DiverseVulPaper to cite for DiverseVul

  • Draper VDISC

    Vulnerable code and software securityFunction-level vulnerability detection

    The Draper Vulnerability Detection in Source Code (VDISC) dataset contains C and C++ functions from open source software. Labels come from the findings of three different static analysis tools.

    Known weaknesses

    Labels come from static analyzer findings rather than confirmed and patched vulnerabilities. Many labeled findings may therefore be false positives that mislead trained models.

    Access
    Open download
    License
    Not stated
    Data type
    HDF5 files with function source and labels
    Size
    1.27 million functions
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for Draper VDISCPaper to cite for Draper VDISC

  • MegaVul

    Vulnerable code and software securityFunction-level vulnerability detection

    MegaVul extracts vulnerable and non-vulnerable functions from CVE fix commits across 28 Git hosting sites. It provides several code representations for training sequence-based and graph-based models.

    Known weaknesses

    The GPL-3.0 license places conditions on redistribution of derived datasets and tools. Java data was added after the paper and has different coverage from the C and C++ data.

    Access
    Open download
    License
    GPL-3.0
    Data type
    JSON functions with code representations
    Size
    17,380 vulnerabilities from 992 repositories across 169 CWE types
    Time span
    2006-01 to 2023-10
    Link checked
    2026-09-27

    Dataset page for MegaVulPaper to cite for MegaVul

  • MoreFixes

    Vulnerable code and software securityVulnerability fix commits

    MoreFixes mines CVE fix commits through improved repository discovery with the Prospector tool. It extends the CVEfixes schema with many more projects and fix commits.

    Known weaknesses

    The authors estimate about 4 percent noise from Prospector heuristics in the final dataset. Use the database dump because rebuilding from scratch takes days or weeks. A 2026 release is larger than the version described in the paper.

    Access
    Open download
    License
    Not stated
    Data type
    Database dump and patch files
    Size
    26,617 CVEs, 6,945 projects, 31,883 fix commits (paper version)
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for MoreFixesPaper to cite for MoreFixes

  • NIST SARD Juliet Test Suites

    Vulnerable code and software securitySynthetic vulnerable test cases

    The NIST Software Assurance Reference Dataset (SARD) hosts the Juliet test suites for static analysis evaluation. Each test case contains flawed and fixed code mapped to a specific CWE.

    Known weaknesses

    Juliet code is synthetic and much simpler than real production software. Models trained only on Juliet often fail to generalize to real open source projects.

    Access
    Open download
    License
    Not stated
    Data type
    Source code archives with CWE labels
    Size
    Juliet C/C++ 1.3 has 64,099 cases, Java 1.3 has 28,881, C# 1.3 has 28,942
    Time span
    2017-2022 releases
    Link checked
    2026-09-27

    Dataset page for NIST SARD Juliet Test Suites

  • PrimeVul

    Vulnerable code and software securityFunction-level vulnerability detection

    PrimeVul is a C and C++ function dataset built with stricter labeling and thorough deduplication. It uses chronological splits and paired vulnerable and patched samples for realistic evaluation.

    Known weaknesses

    The paper shows that models scoring well on older datasets perform poorly on PrimeVul. Labels are more accurate than earlier datasets but still come from automated rules.

    Access
    Open download
    License
    MIT
    Data type
    JSONL functions with commit and CVE metadata
    Size
    About 7,000 vulnerable and 229,000 benign functions covering more than 140 CWEs
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for PrimeVulPaper to cite for PrimeVul

  • Project KB (SAP)

    Vulnerable code and software securityVulnerability fix commits

    Project KB from SAP maps publicly disclosed vulnerabilities in open source Java projects to their fix commits. The knowledge base continues to grow through statements contributed by the community.

    Known weaknesses

    The dataset centers on Java libraries that SAP products or internal tools use. Current counts exceed the figures reported in the 2019 paper and should be recomputed.

    Access
    Open download
    License
    Apache-2.0
    Data type
    YAML statements
    Size
    624 vulnerabilities, 205 Java projects, 1,282 fix commits (2019 paper)
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for Project KB (SAP)Paper to cite for Project KB (SAP)

  • ReposVul

    Vulnerable code and software securityRepository-level vulnerability detection

    ReposVul is a repository-level vulnerability dataset with untangled patches and inter-procedural context. It includes function, file, and repository level data across four programming languages.

    Known weaknesses

    Untangling relies on LLM and static analysis filters that can still mislabel code changes. The repository has no stated license, so reuse terms should be confirmed with the authors.

    Access
    Open download
    License
    Not stated
    Data type
    JSON records with code and patch context
    Size
    6,134 CVEs across 236 CWE types and 1,491 projects
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for ReposVulPaper to cite for ReposVul

  • ReVeal (Chromium and Debian)

    Vulnerable code and software securityFunction-level vulnerability detection

    ReVeal collects vulnerable and fixed C functions from the Chromium and Debian issue trackers. The repository also includes the Devign FFmpeg and QEMU subset used in the same study.

    Known weaknesses

    Data downloads rely on Google Drive links that may change or disappear. The PrimeVul paper reports label noise and duplication across this and similar datasets.

    Access
    Open download
    License
    MIT
    Data type
    Function files with Joern-based processing scripts
    Size
    Not stated
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for ReVeal (Chromium and Debian)Paper to cite for ReVeal (Chromium and Debian)

  • SecurityEval

    Vulnerable code and software securitySecure code generation benchmark

    SecurityEval provides prompts that ask code generation models to complete security-sensitive Python code. Each prompt maps to a CWE and supports static analysis of the generated code.

    Known weaknesses

    The benchmark is small, so results can vary widely between runs and models. Evaluation depends on static analyzers that can miss or misreport security issues.

    Access
    Open download
    License
    Not stated
    Data type
    JSONL prompts
    Size
    121 prompts covering 69 CWEs (v2.1)
    Time span
    Not stated
    Link checked
    2026-09-27

    Dataset page for SecurityEvalPaper to cite for SecurityEval

Descriptions and caveats come from the CCAIR dataset catalog. Each entry shows the date its link was last checked; confirm access and license terms on the dataset's own page before you use it.

Related Pathway Finder roles

Each role guide covers the work, the realistic first job, starter certifications and free practice.

AI Security SpecialistSOC AnalystThreat Intelligence Analyst

Sources

We checked each source in October 2026. Rules, prices and dates change, so confirm them on the source before you act.

  1. [1] MITRE. ATT&CK.
  2. [2] CVE Program. cve.org.
  3. [3] NIST. National Vulnerability Database.

Last verified: October 2026. Written for students by the Center for CyberAI Research (CCAIR) at Georgia State University. Found an outdated fact? Tell us through the contact page.