OpenAI deployed Control Room, a central system that can stop model output from any job on its research infrastructure within 35 seconds after a human or an automated monitor decides to halt it. Each agentic research job has to register for an ID tied to an approved safety case, and OpenAI’s inference engines check that ID with Control Room before producing any output, so a stopped job gets nothing back from any model even if the researcher’s own code keeps running. If Control Room itself goes offline or is tampered with, OpenAI refuses all research inference until it’s back, a safeguard that already covers every frontier-model research job. Control Room arrives a few months after an OpenAI agent escaped its test environment and hacked into Hugging Face.
