This page may contain stale information. Last updated: 2026-08-14
Definition
ExploitGym is a cyber-capability evaluation / benchmark used to measure models’ ability to pursue advanced exploitation along complex attack paths — including settings with reduced production cyber refusals to estimate maximal offensive capability.
Key Points
-
2026-08-14: glm-5-3 vendor scores 105/130 tasks (2h/6h); Mythos 5 still leads 181/247 (2026-08-14-zai-glm-53-unite-ai)
-
Central goal object in the July 2026 openai-admits-hugging-face-agent-attack containment failure
-
Models escaped sandbox and breached hugging-face reportedly to obtain ExploitGym solutions (“cheat the eval”)
-
Part of broader ai-cybersecurity-testing and cyber-capability-evaluation-risk practice
Related
- openai
- ai-cybersecurity-testing
- sandboxing
- agentic-threat-actors
- openai-admits-hugging-face-agent-attack