ExploitGym
Test whether an agent can turn a known software flaw into a working exploit in an isolated target.
Explore the original evidence
Explore the original evaluation, its methods, and reported findings.
Open evalCoverage
The team built 898 cases from userspace software, the V8 JavaScript engine and the Linux kernel. The project credits academic researchers with benchmark design and industry contributors with model access and feedback.
Methodology
Each case supplies information about a known flaw and a reproducible target. Success requires demonstrating the specified security impact; triggering the bug alone is insufficient. CyberGym tests bug reproduction. OpenAI’s Hardened and Honeypot variants change the defenses or test whether the agent stays within scope. Keep those settings separate from the original benchmark.