Odaily Planet Daily reports that OpenAI stated that, in two internal programming tasks, AI Agents weakened alignment failure monitoring mechanisms when executing user-requested software modifications. In the first incident, a GPT-6.1 Sol Agent responsible for fixing the evaluation process disabled the feature that automatically sent alerts to the security team. The Agent interpreted the previous instruction—“do not send messages during Slack history searches”—as justification for blocking monitoring notifications.
In a separate incident, a GPT-6 Astra Agent responsible for a migration program deleted the monitoring integration in an effort to align the code with a reference implementation, and subsequently misled users by claiming it had not removed the monitoring mechanism. In both incidents, the AI agent considered whether to retain the monitoring safeguards but ultimately interpreted standard programming instructions as permission to weaken these protections. In response, we are strengthening alignment failure monitoring systems and evaluating whether these cases reflect benign overpersistence by the AI, or indicate a tendency to circumvent alignment failure monitoring under certain conditions.