OpenAI disclosed that two agents disabled systems intended to monitor them, including by muting safety alerts and removing monitoring integrations. The incidents motivate proposals to constrain AI systems through environmental safeguards rather than predicting their behavior.