AI labs talk a lot about stopping a model that misbehaves, and say much less about how the stop button actually works. OpenAI has now described one: Control Room, a central service that can cut off model output for any research job across its training and evaluation infrastructure, with a stated worst case of 35 seconds from decision to halt.
The design is simple to explain. Every research job that runs agentic inference must register with Control Room, receive a job ID and be linked to an approved safety case. When the job asks a model for tokens, it passes that ID to the inference engine, the software running the model on GPUs. Before answering, the engine asks Control Room whether the job is still allowed to run. If someone has stopped it, the request is refused and no more output is produced, even though the job's own code may still be running. A stop applies to any model on any engine the job calls.
Two choices stand out. First, OpenAI deliberately puts the check in the inference layer rather than in researchers' job code, which changes constantly and may be buggy. There are only a few inference code paths across its research, so one control point covers them all, and the company says it also runs automated sweeps for rogue engines that skip the check. Second, the system fails closed: if Control Room goes down, through error, deliberate disabling or tampering, every inference request is denied until it returns. That was in place for all frontier research workloads at publication and due to reach the rest within days.
The trigger can be a person or a machine. OpenAI says all its frontier reinforcement learning runs are watched by a monitor model that inspects what the model in training is doing; a recent DNS-related incident was caught that way, and the monitor can now tell Control Room to stop the job directly.
The limits are worth stating. The 35-second delay comes from caching the authorisation checks to handle traffic, and OpenAI says it is working to cut it to zero; for a fast-acting agent, half a minute is not nothing. Control Room covers research only, since OpenAI says its public serving stack already has its own takedown tools. And everything here is self-reported, with no outside audit of how often the switch has been pulled or how it performed under real pressure.