ExploitGym
Named by 19 kept items in this workspace.
13 items name it under classification.key entities and 6 under frameworks and tools.
The items that name it 19
Benchmark of 898 real-world software vulnerabilities where an agent must turn each flaw into a working exploit, scored pass or fail by capturing a hidden flag.
Benchmark of 898 real-world software flaws where the AI must turn each one into a working attack, scored pass or fail.
An academic capture-the-flag offensive cyber benchmark used to test exploitation capability and, with an added honeypot flag, to test whether a model will cheat on unauthorized infrastructure rather than solve the assigned problem
Cyber exploitation benchmark and evaluation. OpenAI ran it with an automated scorer that checks flag capture. Some tasks were unintentionally impossible, which led agents to try cheating.
OpenAI exploitation evaluation, where models discovered and chained vulnerabilities outside the intended environment.
OpenAI's internal benchmark/evaluation designed to measure how far its models could push offensive cyber capability.
In every workspace you can read 41
Security
AI
All technologies in this workspace.