Verified86%Harbor: pendingSafety & AlignmentCoding

Cybench

Capture-The-Flag Cybersecurity Challenges

Tests language models on cybersecurity skills using 39 of 40 practical, professional-level challenges taken from cybersecurity competitions, designed to cover various difficulty levels and security concepts.

How to run

ltb · cybench
$ npm i -g lettokenburn-cli
$ ltb bench pull cybench

External harness. pip install inspect-evals && inspect eval inspect_evals/cybench --model <provider/model> https://github.com/UKGovernmentBEIS/inspect_evals

Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs

Scores

23 sourced entries
HarnessSourceConfidenceLink
Claude Opus 4.693.0N/ASelf-reportedmediumN/Awww-cdn.anthropic.com
Claude Opus 4.5 (2025-11-01)82.0N/ASelf-reportedmediumN/Aassets.anthropic.com
Claude Sonnet 4.5 (2025-09-29)60.0N/ASelf-reportedmediumN/Aassets.anthropic.com
Claude Sonnet 4.5 (2025-09-29)55.0N/AIndependentmediumN/Acybench.github.io
Grok 4 070943.0N/ASelf-reportedmediumN/Adata.x.ai
Claude Opus 4.1 (2025-08-05)42.0N/AIndependentmediumN/Acybench.github.io
Grok 4.139.0N/ASelf-reportedmediumN/Adata.x.ai
Claude Opus 4 (2025-05-14)38.0N/AIndependentmediumN/Acybench.github.io
Claude Sonnet 4 (2025-05-14)35.0N/AIndependentmediumN/Acybench.github.io
Grok 4 Fast30.0N/ASelf-reportedmediumN/Adata.x.ai
o3-mini22.5N/AIndependentmediumN/Acybench.github.io
Claude 3.7 Sonnet Thinking (2025-02-19)20.0N/AIndependentmediumN/Acybench.github.io
Claude 3.5 Sonnet17.5N/AIndependentmediumAug 15, 2024arxiv.org
Claude 3.5 Sonnet17.5N/AIndependentmediumN/Acybench.github.io
GPT-4.5 Preview (2025-02-27)17.5N/AIndependentmediumN/Acybench.github.io
GPT-4o12.5N/AIndependentmediumN/Acybench.github.io
Claude 3 Opus10.0N/AIndependentmediumN/Acybench.github.io
o110.0N/AIndependentmediumN/Acybench.github.io
o1-mini10.0N/AIndependentmediumN/Acybench.github.io
Gemini 1.5 Pro 001 Feb247.5N/AIndependentmediumN/Acybench.github.io
Llama-3.1-405B-Instruct7.5N/AIndependentmediumN/Acybench.github.io
Mixtral-8x22B-Instruct-v0.17.5N/AIndependentmediumN/Acybench.github.io
Meta-Llama-3-70B-Instruct5.0N/AIndependentmediumN/Acybench.github.io

Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.

Reproduced by the community

All runs
Loading community runs…

About

Tests language models on cybersecurity skills using 39 of 40 practical, professional-level challenges taken from cybersecurity competitions, designed to cover various difficulty levels and security concepts. The motp challenge is excluded due to GPL licensing.