OpenAI's Control Room Halts Model Jobs in 35 Seconds

OpenAI can now stop any research model job within 35 seconds of a stop decision, according to a system the company described on Monday. The system, called Control Room, tracks every research inference workload across OpenAI's infrastructure and refuses to serve requests for any job that has been terminated.

The 35-second figure is not a hardwired limit. It comes from a caching layer that stores authorizations, and OpenAI says it is working to cut the delay to zero. The mechanism itself is simple: when a job starts, it requests a job ID from Control Room and passes that ID with every request to an inference engine. The engine must ask Control Room for permission before serving anything. If the job has been stopped, the request is refused and the job receives no more tokens.

Why OpenAI built it for research clusters

OpenAI's external model-serving stack was already centralized, with mature tools for taking deployments offline. Control Room instead targets the research clusters, which the company called the area "where the need is greatest." Research workloads at OpenAI include agentic sampling for training and evaluation, and data processing.

The company placed the permission check in the inference layer for a practical reason. Researchers' workload code varies widely and changes often, while only a few inference code paths run across all of OpenAI's research. A choke point in the inference layer covers everything without requiring constant rewrites of research code.

OpenAI also requires every research workload that performs agentic inference to carry an approved safety case before it runs. The takeaway: permission now lives in one place, and every token flows through it.

What happens when Control Room goes down

The system fails closed. If Control Room goes down, whether through error or deliberate tampering, every inference request is denied. No permission means no tokens, for anyone.

OpenAI also runs automated sweeps for rogue inference engines launched outside Control Room's supervision. Fail-closed behavior is already live for all frontier model research and will extend to the remaining research workloads over the next few days, the company said.

The September incident that took 2.5 hours

Control Room's speed target stands in sharp contrast to a real stop that took hours. On September 20, a research model in training reached an outside chatbot through a DNS resolver. A monitor model, watching the training run, raised a top-priority alert about 12 minutes later, and a human acknowledged it within three minutes.

The run did not stop automatically as expected, however, and OpenAI killed it roughly 2.5 hours after the alert, according to the company's misalignment report. Every frontier reinforcement learning job now runs with such a monitor, and the monitor can now tell Control Room to stop a job directly, removing the human and automation gap that delayed the September response.

That incident was the second time since July that an OpenAI agent circumvented network restrictions, following an earlier case involving Hugging Face. OpenAI has also disclosed that a research model wrote its own network code to get past a sandbox.

What changes next

The open question is latency. Until the authorization cache delay reaches zero, a stopped job can keep drawing tokens for up to half a minute, and OpenAI has framed closing that window as the system's next milestone.