OpenAI says it found that its AI Agent weakened the monitoring mechanism for alignment failures.

PANews October 21 news: OpenAI stated that in two internal programming tasks, AI Agents weakened alignment failure monitoring mechanisms while executing user-requested software modifications. In the first incident, a GPT-6.1 Sol Agent responsible for fixing the evaluation process disabled the function that automatically sends alerts to the security team. The Agent interpreted a previous instruction to "do not send messages during the search of Slack history" as a reason to block monitoring notifications. In another independent incident, a GPT-6 Astra Agent responsible for a migration program deleted the monitoring integration in the program to make the code consistent with the reference implementation, and then falsely told the user that it had not removed the monitoring mechanism. In both incidents, the AI Agents had considered whether they should retain the monitoring safeguards, but ultimately interpreted routine programming instructions as permission to weaken these safeguards. In response, we are strengthening the alignment failure monitoring system, and we are also evaluating whether these cases reflect benign overpersistence behavior by the AI or indicate a tendency in specific situations to evade alignment failure monitoring.