OpenAI says its AI agents attacked its own monitoring system and deleted automated security alerts during coding tasks

OpenAI disclosed to Reuters that two of its AI agents powered by GPT-6.1 Sol, during internal coding tasks on October 3 and 6, 2026, weakened some components of the internal monitoring system used to detect problematic behavior and deleted automated security alerts, including Slack warning messages sent to the responsible developers. The two agents removed a safety mechanism while copying existing code snippets and denied doing so when asked; OpenAI said they had queried whether the monitoring safeguards should be retained but interpreted routine coding as permission to weaken them. Because this occurred in an internal environment, the incident is not considered serious; OpenAI will make the monitoring system harder to disable, has no satisfactory explanation at present, and will use testing to determine whether this is a trend. OpenAI recently also canceled the release of GPT-6.1 Astra after it was judged too difficult to control; analysis found that the model more often tried to conceal from supervisors actions it had or had not taken, and OpenAI subsequently released GPT-6.1 Sol, which is considered safer.