CyberGym

1,507 real patched vulnerabilities across 188 projects; agents must write a PoC that crashes the unpatched binary but not the fix.

Also known as: CyberGym: Evaluating AI Agents' Real-World Cybersecurity Capabilities at Scale

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
Subcategoryproof-of-concept generation for historical OSS-Fuzz/ARVO vulnerabilities
Page statusactive
Metricreproduction rate (PoC crashes unpatched and not patched); new-vulnerability rate (PoC also crashes the patch)
Directionhigher_is_better
Unit%
Dataset size1507
Dataset licenceApache-2.0 (official repository); bundled project trees keep their upstream licences
PublisherUniversity of California, Berkeley

What it measures

CyberGym asks an agent to reproduce a historical memory-safety bug. The default (level-1) setting gives the unpatched codebase plus a short vulnerability description. The agent must emit a proof-of-concept file that crashes the pre-patch build under sanitizers and does not crash the post-patch build. A PoC that also crashes the patched program is counted as a new vulnerability, not a reproduction. This is live tool use against compiled targets, not a quiz about CVEs.

Task format

Docker sandbox with a solver container and separate vulnerable and fixed program containers. The agent writes a PoC and submits it through submit.sh. inspect_evals defaults to a react solver, max_attempts=3, level1 only, no k8s. Four difficulty levels change which files are mounted (code only, plus description, plus crash log, plus patch and fixed tree).

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub