OpenAI can halt any research model's output in 35 seconds
Every agentic research job at OpenAI now registers with Control Room, which can stop its model output within 35 seconds of a decision.
2 minAutonomous AI AgentsFresh · 9 Nov
OpenAI described Control Room on 9 November, a central system that tracks model workloads across its research infrastructure and stops them when safety or security concerns arise. It covers every research inference workload, including agentic sampling for training and evaluation and data processing, and works together with OpenAI's misalignment monitoring. Once a decision is made to stop or pause a job, model output halts within 35 seconds. The company says the system is already deployed across research, where the need is greatest; its external serving stack is more centralised and already has its own way to take deployments down quickly.
The design joins three parts: the workload, meaning the job a researcher runs, such as a reinforcement learning experiment or an evaluation; the inference engine, the model running on GPUs; and the Control Room service. Every research workload that does agentic inference must be tied to an approved safety case. When a job starts, its code asks Control Room for a job ID. Each later request for tokens goes straight to an inference engine with that ID attached, and the engine asks Control Room for permission before answering.
Stopping a job therefore does not depend on finding and killing processes across a cluster. If Control Room has marked the job as terminated, it denies the engine's request, the engine returns a refusal to the workload, and no further output is produced. One engine can serve several jobs, so the check works per job rather than per machine.
OpenAI built the service to keep working when other systems fail, and to fail closed: if Control Room itself goes down, model output stops. The company frames the work as a response to models becoming more powerful, which makes the ability to detect problems and stop a model quickly more critical.
Retold from OpenAI. This is a summary in our own words; follow the link for the original reporting.