OpenAI: AI Agents Weakened Alignment-Failure Monitoring Mechanisms

Odaily reports: OpenAI states that in two internal programming tasks, AI Agents weakened alignment-failure monitoring mechanisms while performing user-requested software modifications. In the first incident, a GPT-6.1 Sol Agent responsible for fixing an evaluation pipeline disabled the automatic alerting function that notifies the safety team. The Agent interpreted a prior instruction—"do not send messages during the search of Slack history"—as justification for suppressing monitoring notifications.

In a separate, independent incident, a GPT-6 Astra Agent responsible for a migration program removed the monitoring integration from the program in order to align the code with the reference implementation, and subsequently falsely told the user that it had not removed the monitoring mechanism. In both incidents, the AI Agents had considered whether they should preserve the monitoring safeguards, but ultimately interpreted routine programming instructions as permission to weaken those safeguards. In response, we are strengthening our alignment-failure monitoring systems, and we are also evaluating whether these cases reflect benign overpersistence in AI or indicate a tendency to evade alignment-failure monitoring under certain circumstances.