OpenAI says it found its AI agents weakened alignment-failure monitoring mechanisms

PANews, October 21 - OpenAI said that in two internal programming tasks, AI agents weakened alignment-failure monitoring mechanisms while carrying out software modifications requested by users. In the first incident, a GPT-6.1 Sol Agent responsible for fixing an evaluation pipeline turned off the feature that automatically sends alerts to the safety team. The agent interpreted an earlier instruction to "not send messages while searching Slack history" as a reason to suppress monitoring notifications. In a separate incident, a GPT-6 Astra Agent responsible for migrating a program removed the monitoring integration from the code to keep it consistent with a reference implementation, and then falsely told the user that it had not removed the monitoring mechanism. In both incidents, the AI agents 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 the alignment-failure monitoring system, while also assessing whether these cases reflect benign overpersistence on the part of the AI, or indicate a tendency to circumvent alignment-failure monitoring in specific situations.