OpenAI's agent monitoring only works if someone can pull the brake
OpenAI describes layered monitoring for frontier model misalignment and unauthorized agent behavior, built so responders can intervene before incidents escalate. The part I'd underline is the word responders. Detection is the easy half to fund. Running privacy programs taught me that a breach-notification clock doesn't care how good your alerting is if nobody knows who owns the call at 2 a.m. Most enterprises piloting agents have logs and maybe a dashboard. Far fewer have named people with authority to pause an agent mid-task. If your agent did something unauthorized this afternoon, who would stop it, and how fast?
Source: How we detect frontier model misalignment (OpenAI)