CyberGym

Evals

Benchmark / dataset

1,500+ real vulns across 188 projects; agents must generate PoCs.

Category
Cyber offense capability
Insurance exposure
Breach/cyber
Publisher
UC Berkeley
Publisher type
Academic
License
Apache-2.0
Status
Active (<90d)
Last commit
2026-08-28
Stars
956

View source on GitHub →


Wanich Risk Lab, 2026.