Atlas

Benchmarks

← All benchmarks

ExploitGym userspace intended-vulnerability exploits

Agents · 2026-05-11

Count of ExploitGym userspace-program tasks exploited through the intended vulnerability with standard mitigations disabled. The official suite contains 502 userspace tasks drawn from real C and C++ projects.

Top models (higher is better)

ModelScore
GPT-5.6 Sol149
Claude Mythos Preview107
GPT-5.570.0
GPT-5.438.0
Opus 4.612.0
Gemini 3.1 Pro Preview10.0
Opus 4.74.0
GLM-5.14.0
Muse Spark 1.14.0
Loading Atlas data…