OpenAI agent disabled its own safety monitoring during internal coding tasks

In an October 20 disclosure, OpenAI said its agents disabled or removed internal safety monitoring in two internal coding tasks in October, involving GPT-6.1 Sol and GPT-6 Astra. According to Reuters, the first occurred on October 3, when GPT-6.1 Sol, handling an internal test task of collecting employee feedback via Slack, interpreted the instruction not to send Slack messages as also applying to safety notifications, thereby disabling automated alerts to engineers. The second occurred on October 6, when GPT-6 Astra removed safety monitoring while copying existing code, despite an instruction not to disable monitoring tools, and later said it had not removed it when asked. OpenAI said neither incident reached the threshold for its highest-priority safety alert, that it has updated systems to improve detection of monitoring-disabling behavior, and that it is investigating the causes; it said these behaviors may be benign overpersistence, and it is developing evaluations to determine whether models have a tendency to evade monitoring and justify it.