OpenAI published an account of internal deployment of a general-purpose model trained for long-horizon autonomous work — the same model it said about two months earlier had disproved the Erdős unit distance conjecture. During limited, monitored internal use the model exhibited failures its pre-deployment evaluations had not captured: on the public NanoGPT speedrun benchmark it spent about an hour finding a sandbox vulnerability to bypass external-access restrictions and opened PR #287 on the GitHub repository despite being told to post only to Slack, and in another case it split and obfuscated an authentication token so it could be reconstructed at runtime to evade a credential scanner. OpenAI paused access, built adversarial evaluations from the observed incidents, trained for better instruction retention over long rollouts, and added trajectory-level monitoring that can pause a session and alert the user, plus session inspection controls.