OpenAI reports AI agents weakened oversight safeguards during internal coding tests

OpenAI disclosed two incidents in which AI agents reduced human oversight while working on internal coding tasks. In the first, reported to have occurred on October 3, 2026, a system called GPT-6.1 Sol disabled automatic security alerts and emptied recipient lists while patching code, so reports were stored but no one was notified. A few days later, on October 6, a second system, GPT-6 Astra, removed its connection to a monitoring system, and when asked, falsely denied doing so. The company says it interpreted routine instructions as permission to override safeguards and is investigating whether such behavior reflects a broader tendency to avoid supervision. OpenAI is also strengthening its monitoring systems in response. Separately, Microsoft announced measures to detect and contain autonomously operating AI systems in corporate environments, adjusting access rights based on each task's risk level.