Cybench

Tests whether an autonomous LLM agent can solve 40 professional-level capture-the-flag cybersecurity tasks in a live sandbox, with 17 of them broken into guided subtasks for partial credit.

Also known as: Cybench: A Framework for Evaluating Cybersecurity Capabilities and Risks of Language Models

unassessed

This page is a discovery lead. Nobody has yet assessed it against the catalogue contract, so it carries no disposition. Absence of evidence here is not evidence of staleness.
Categoryagentic
Subcategoryautonomous capture-the-flag (CTF) cybersecurity task-solving, with a graded subtask mode
Page statusactive
Metrictask success rate (fraction of the 40 CTF tasks solved), reported separately per scoring mode
Directionhigher_is_better
Unit%
Dataset size40
Dataset licenceApache-2.0 (repository code); the redistributed CTF challenge content itself is not uniformly licensed -- see Dataset and licence
PublisherStanford University (Center for AI Safety collaborators Dan Boneh, Daniel E. Ho and Percy Liang among the senior authors), with a 27-author team

What it measures

Cybench measures whether an autonomous LLM agent, given shell and Python tool access inside a Kali Linux sandbox, can solve real capture-the-flag (CTF) security challenges drawn from four professional-level competitions (HackTheBox Cyber Apocalypse 2024, SekaiCTF, Glacier and HKCert). Tasks span six domains -- cryptography, web security, reverse engineering, forensics, exploitation (pwn) and miscellaneous -- and the paper anchors difficulty in human terms: the times competitive human teams took to first-solve each task ranged from a few minutes up to nearly 25 hours. This page categorizes Cybench as agentic rather than domain knowledge, because success depends on multi-step planning and tool use against a live target, not a static answer.

Task format

An agent is placed in a Kali Linux Docker sandbox with bash and code-execution tools and must find and submit a flag string for each of 40 tasks. 17 of the 40 tasks additionally ship a sequence of subtasks -- guided intermediate questions with their own answers -- scored sequentially for gradated credit. Three scoring modes result: unguided performance (binary success on the full task with no hints), subtask-guided performance (binary success on only the final subtask, i.e. full guidance up to the last step), and subtask performance (the fraction of a task's subtasks solved).

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub