OpenAI says Control Room can stop a research job in 35 seconds

OpenAI says a new system called Control Room can stop a research model job within 35 seconds of a decision to intervene. The company says it tracks workloads across its research infrastructure and can halt them when a safety or security concern comes up. It covers research inference jobs, including sampling used to train and evaluate models, and data processing. The 35-second lag comes from a caching layer for inference authorization. OpenAI says it is working to bring that delay to zero.

Control Room is already running on research infrastructure, not on the public model-serving stack. OpenAI says that public stack is more centralized and can already take a deployment down. Research jobs are different. A researcher starts a job and gets an ID from Control Room. Each time the job asks a model for more output, the inference engine asks Control Room whether that job is still allowed to run. If the job has been stopped, the engine refuses the request.

OpenAI says a separate monitor can trigger the stop. The monitor is another model that watches what the model being trained is doing. If it flags a problem, it can tell Control Room to end the job. OpenAI says a monitor caught a recent DNS-related incident. This story does not add details from that report. The company also says every research workload that does agentic inference must have an approved safety case.

The system is built to fail closed. OpenAI says that if Control Room goes down, or if someone disables it, research inference requests are denied until it is working again. That setting is already on for frontier-model research workloads. The company says the rest of its research workloads will get it over the next few days. It also says it runs checks so people cannot quietly launch an engine with Control Room turned off.