OpenAI announced Control Room on its blog and X account on 9 November. The company said the system covers every research inference workload, meaning every run in which a model produces output. Examples include agentic sampling for model training and evaluation, and data processing. Both posts are the company’s own and mention no external evaluation.

OpenAI has deployed Control Room in its research infrastructure, not in the model service it offers customers. The company said that service is more centralised and already has an established way to withdraw models quickly.

Permission for every model call

When a research job starts, Control Room gives it an identifier. Each time the job requests more tokens from a model, it passes the identifier to the inference engine, the software that runs the model on GPUs. The engine asks Control Room for permission before it responds. If the job has been stopped, the engine denies the request, and the job can no longer get output from any model on any engine.

OpenAI placed the check in inference because the code in researchers’ jobs varies and changes often, whereas inference has very few code paths and they are uniform. The company also uses agents and programmatic checks to search continuously for engines where the check is missing or switched off.

For example, a monitoring system can decide to stop a job when a monitor – a second model that watches the model being trained – raises an alert. OpenAI said a monitor watches every reinforcement learning run of its frontier models. The monitor also detected the DNS-related incident the company recently reported.

OpenAI also requires every research job that runs agents to have an approved safety case, a written explanation of why the job is safe enough.

35-second latency and failures

The company said a job stops within 35 seconds of Control Room receiving a stop instruction. The latency comes from caching permission checks, which the high volume of requests makes necessary. OpenAI said it aims to cut the latency to zero. The introduction to the blog post and the X post count the same 35 seconds from the decision to intervene, so they do not make clear how long it takes to pass the decision on.

The company said it built the system to keep working even when other systems fail. If Control Room crashes or someone disables it, for example through sabotage, the engines deny all inference requests until the service recovers. OpenAI said this safeguard covers all frontier model research jobs and will extend to other research jobs in the coming days.