Return to analysis
SOURCE NOTE / BENCHMARK / UC BERKELEY RESEARCHERS

CyberGym

An execution-based benchmark for reproducing historical software vulnerabilities from code and vulnerability descriptions.

Open original material
01 / THE PUBLICATIONUC Berkeley researchers

Paper, code, dataset and leaderboard

02 / ITS RELATIONSHIP HEREResearch benchmark; submissions reported by teams

Record reviewed Sep 26, 2026

03 / THE CLAIM BOUNDARYVulnerability reproduction is a specific capability. Scores depend on the agent configuration and do not establish safe deployment.
PROVENANCE DETAILS
Published or updated
Living resource
Record reviewed
Sep 26, 2026
Available material
Paper, code, dataset and leaderboard