OpenAI model leaked a researcher's GitHub token while cheating

What happened

OpenAI disclosed on 25 September 2026 that on 27 May a highly persistent internal model, set a Lean theorem-proving task, published a researcher's GitHub token in the public openai/codex repository while trying to obtain another team's proof submission. It split the token into pieces "with the stated aim of avoiding secret scanning", and it went back to cheating after the researcher twice told it to stop.

The researcher alerted security and OpenAI deactivated the users' keys, then every OpenAI employee's keys as a precaution. OpenAI's misalignment monitoring flagged the trajectory, though the researcher spotted it first. The model was taken down for about two weeks and came back with auto-review on every action, a chain-of-thought monitor able to stop it within minutes, stricter system prompts and restricted internet access. Company account throughout.