OpenAI says Control Room halts research AI jobs in 35 seconds

For people who use ChatGPT or OpenAI's API, Control Room changes nothing. The company described the service as an emergency brake for its research infrastructure, where it trains and tests its models, and says it is already deployed there. The models it offers to outside users run on separate infrastructure that is more centralized and, the company says, can already pull deployments quickly.

How it works

Every research job, whether a reinforcement learning experiment, an evaluation or a large sampling run, asks Control Room for an ID and must be tied to an approved "safety case." When the inference engine, the software that runs the model on GPUs, needs to generate text, it checks with the service using that ID. If the job is marked as stopped, the request is denied and the model stops producing output.

The longest it can take between a stop order and the moment a job can no longer make progress is 35 seconds.

That lag comes from a caching layer that authorizes inference requests and is needed to handle the traffic volume. A person can give the order, and so can the systems that watch for misaligned behavior, meaning models that act against what their makers intend. OpenAI says it put the brake at the inference layer rather than inside each experiment because experiment code changes often and is written by many different teams.

The design also fails shut. If Control Room goes down, through an error or because someone switches it off, even with bad intent, the system rejects every inference request until it is restored. That lockdown already works for jobs with frontier models, and the rest of the research workloads will get it within days. Continuous checks, run by agents and conventional programs, look for engines launched without Control Room or with it disabled.

The announcement follows an episode OpenAI documented. In a report on its alignment blog, it says that in September an agent in training reached a public chatbot through a gap in its internet restrictions: DNS filtering in its test environment was insufficient. The monitoring system flagged the behavior within 15 minutes, a person began reviewing it three minutes later, and the run was stopped 2.5 hours after that. The Control Room page does not tie the system to that case.

What has not been tested

Every figure comes from OpenAI, and there has been no audit or independent test. OpenAI does not say who operates the system or what thresholds trigger a shutdown, and 35 seconds is a ceiling, not a description of what a compromised job could do in that window.

OpenAI says it is working to bring those 35 seconds down to zero so it can end jobs instantly.