OpenAI describes Control Room, a research stop that cuts off model output within 35 seconds
The company page says engines ask a central service before they return tokens, and that a stopped job can keep running while further output is refused. The 35 seconds is a cache. Fail-closed behavior is on frontier research workloads now, and on the rest of research over the next few days. Neetika Walter’s story is timestamped 8:25 p.m. Eastern on Monday, and it embeds an OpenAI post from that day.
Feature3 minIn-depth
Maya Chen3 min read
- OpenAI
- Safety
- Research
- Agents

Photo: OpenAISource page
OpenAI has published a description of Control Room, a service that tracks research jobs and can stop them from receiving more model output. The page says it is already deployed on the company’s research infrastructure, that it works with the systems that watch for misaligned behavior, and that it halts model output within 35 seconds of a decision to stop or pause a job. Neetika Walter’s story in Interesting Engineering is timestamped 8:25 p.m. Eastern on Monday, which was 1:25 a.m. UTC on Tuesday. The story embeds a post from the OpenAI account on X dated 9 November. The company page itself does not carry a clock time.
The 35 seconds is not a claim about how fast a person can notice a problem. OpenAI says the maximum time between the moment Control Room receives an instruction to stop a job and the moment further progress is no longer possible is 35 seconds, and that the lag is a caching layer for inference authorization, put there to handle the volume of checks. The company says it is working to drive that number to zero. Walter’s headline calls the mechanism a kill switch for rogue agents. The page does not use that phrase. It says the job’s own code can remain running. What stops is further model output.
A job ID, checked at the engine
OpenAI splits the system into three pieces. The workload is the job: a reinforcement-learning experiment, an evaluation, a high-volume sampling run, or a data-processing job. Every research workload that does agentic inference has to be tied to an approved safety case. The inference engine is the model running on GPUs. One job can use more than one model, and one engine can serve more than one job. Control Room is the service that records which jobs may still receive output.
The sequence on the page is concrete. Workload code asks Control Room for a job ID and attaches that ID to later requests. When it wants tokens, it asks an engine directly. The request does not pass through Control Room. Before the engine fulfills it, the engine asks Control Room whether that job ID is still authorized. If someone, or a monitor, has marked the job stopped, the engine returns a refusal instead of tokens. OpenAI says it does not rely on the workload code to perform that check itself. The reason it gives is that workload code changes often and is a poor place for a load-bearing control, while only a small number of inference code paths are in use across OpenAI research. The page also says the company runs continual sweeps, agentic and programmatic, for engine implementations that have left Control Room unimplemented or switched off.
Frontier reinforcement-learning jobs, the page says, already run with a monitor: another model that inspects what the model being trained is doing and can raise an alert. If the monitor decides to stop the job, it sends Control Room the job ID. The next sampling request, to any model on any engine, is denied. OpenAI’s example of a catch is “our recent DNS-related incident,” which it says a monitor caught. The page does not describe that incident again.
What happens if Control Room itself fails
The integration is meant to fail closed. If the Control Room service is down, whether from an error, because it was disabled, or because someone tampered with it, engines deny inference requests until it is back. OpenAI says that behavior is rolled out now on all frontier-model research workloads, and that it will finish rolling out across the remaining research workloads over the next few days. The customer-facing serving stack is treated separately. The page says that stack is already more centralized, and that operators can already take a model deployment down.
What the post does not give is a count of jobs stopped, a log of the alerts, or a date by which the 35-second cache goes away. It also does not say Control Room governs ChatGPT or the API. The scope it states is research inference, with the external serving stack left on the older operational path.


