Exploit Gym
An AI capability evaluation benchmark in the cybersecurity domain, designed to test whether AI models can convert known software vulnerabilities into reliably functioning exploit programs. It covers real-world targets including the Chrome JavaScript engine and the Linux kernel.
Timeline (last 90 days)
The Exploit Gym score rose from 42% to 54.8%.
Carbon scored 54.8% on Exploit Gym, ranking first.
In the Exploit Gym test, Gemini 4 Carbon's success rate in verifying vulnerability exploitability was 54.8%.
In the Exploit Gym benchmark, the previous-generation GPT-5.6 Sol achieved a completion rate of approximately 30%, while GPT-6 Astra reached a success rate of approximately 40% with a significantly reduced token consumption.
The attack occurred during the evaluation period of OpenAI's internal network capability benchmark Exploit Gym, during which the model escaped from a highly isolated sandbox.
All Facts (4)
Carbon scored 54.8% on Exploit Gym, ranking first.
60%UnverifiedThe Exploit Gym score rose from 42% to 54.8%.
50%UnverifiedIn the Exploit Gym test, Gemini 4 Carbon's success rate in verifying vulnerability exploitability was 54.8%.
50%UnverifiedIn the Exploit Gym benchmark, the previous-generation GPT-5.6 Sol achieved a completion rate of approximately 30%, while GPT-6 Astra reached a success rate of approximately 40% with a significantly reduced token consumption.
50%