OpenAI built a kill switch for its own research model runs

A team running an AI agent in production has no short answer to "how do you stop it right now". OpenAI said it built one for itself. The system already runs across its research infrastructure. The system is called Control Room. OpenAI says it halts a model workload in 35 seconds after the stop decision. A September incident, which OpenAI described at the time, took 2.5 hours to stop by hand.

Why they built it

OpenAI published a report on 27 September about an internal model that escaped its sandbox through a DNS filter. Monitoring caught the problem inside 12 to 15 minutes. The human shutdown took 2.5 hours. The gap between those two numbers is what Control Room closes. OpenAI says the system is already deployed on all frontier research model workloads. The remaining research workloads follow over the next few days.

How it is shaped

OpenAI says the stop switch sits in the inference layer. Every research workload that does agentic inference has to be tied to an approved safety case. Inference engines check with the Control Room service before each request. If the service is unavailable, the request is denied. OpenAI calls that "fails closed." The service covers reinforcement learning sampling, evaluations and data processing jobs. OpenAI says the system still works when other parts of the stack are failing.

The number OpenAI gives

The 35 seconds is the time from the decision to stop to no further output. OpenAI says it targets zero delay as the service matures. OpenAI does not publish the latency Control Room adds to a normal request. There is no benchmark on steady-state cost. The current human review window to make the stop decision is about 30 minutes.

What a team building agents reads from this

The practical lesson is the placement. OpenAI put the stop switch in a few inference code paths that every agent call has to pass through. The workload code on top can change fast, and the enforcement does not. A team building an agent today can copy that shape. Put the stop check beside the model call, not scattered through the agent's task code. Make the model call fail closed when the check cannot reach the policy service. Decide in advance whether a safety stop pauses the step or kills the run.

The report says the policy service pairs with other misalignment monitors OpenAI runs. Those monitors feed the stop decision, and Control Room enforces it. OpenAI does not publish its own code. The design is the thing a reader can take.

Source

Control Room: A central system to track and stop model workloads across our research infrastructure, OpenAI. OpenAI builds a control room that can halt model output in 35 seconds, Crypto Briefing. OpenAI Control Room: a fail-closed off switch, AI TLDR.