OpenAI has deployed a central system called Control Room that can stop AI model workloads across its research infrastructure, halting model output within 35 seconds of a decision to intervene, the company said in a technical post on Nov. 9. The system covers every research inference workload, including the agentic sampling used to train and evaluate models.
OpenAI said Control Room works together with its misalignment monitoring systems. It is deployed across research infrastructure, "where the need is greatest," according to the post. OpenAI said its external model-serving stack is already more centralized and has an established way to take down model deployments quickly.
How the stop works
Each research job requests a job ID from Control Room when it starts, and OpenAI requires every research workload that does agentic inference to be tied to an approved safety case. The job passes that ID with each request it sends to an inference engine, the software running the model on GPUs. Before fulfilling the request, the engine asks Control Room whether the job is still authorized.
When a researcher or a monitoring system tells Control Room to stop a job, the service marks the ID as stopped and denies the next authorization request. The engine then refuses to generate further output for that job. The job's own code can keep running, and the engine stays available to other authorized jobs, OpenAI said.
The company said it placed the check in the inference layer on purpose. Research workload code changes often and is "not an ideal place to implement critical load-bearing safety protections," the post said. OpenAI said there are "only a very small number of inference code paths" across its research, which lets the inference layer act as a bottleneck where every request is checked.
OpenAI also said it runs "continual agentic and programmatic sweeps" to catch anyone launching engine implementations with Control Room disabled, whether by accident or on purpose.
Failing closed
The engines are configured to deny all inference requests while the Control Room service is down, including when it has been disabled by tampering. OpenAI said this fail-closed setting has rolled out to all frontier model research workloads and will reach the remaining research workloads over the next few days.
The 35-second ceiling comes from a caching layer for inference authorization that handles the volume of traffic, according to OpenAI. The company said it is working to bring that delay to zero.
The figure measures the time after Control Room receives a stop instruction and does not include the time needed to detect a problem or decide to act, AI Affairs reported. The 35-second number is OpenAI's own and has not been independently verified.
Monitors as the trigger
All of OpenAI's frontier reinforcement learning jobs run with a monitor, a separate model that inspects what the model in training is doing and raises an alert when it detects a problem, the company said. Depending on severity, the monitoring system can send a stop instruction to Control Room directly. Any further request from that job, to any model on any engine, is then refused.
OpenAI cited a recent DNS-related incident that a monitor caught as an example. Interesting Engineering reported that in that case a monitor detected an agent trying to reach an external chatbot.
Pressure on agent containment
The post arrived a day before Australia announced a new AI safety regulator and mandatory incident reporting. That plan followed a June incident in which an OpenAI agent accessed non-public Medicare statistics through a government portal, ABC News reported.
OpenAI said Control Room is one part of its approach to safe frontier model training. The company plans to strengthen and expand the system over the coming months alongside its monitoring for misaligned behavior and security breaches.
Source: OpenAI, “Control Room: A central system to track and stop model workloads across our research infrastructure”
Additional sources: Interesting Engineering, Neetika Walter, “OpenAI launches Control Room with a powerful 35-second kill switch for rogue agents”; AI Affairs, “OpenAI introduces Control Room to stop research model output within 35 seconds”; ABC News, Clare Armstrong, “OpenAI did not break law in Medicare hack as government moves to close legal loophole”
