Control Room
OpenAI deployed Control Room, a central system that tracks research inference workloads and can stop their model output within 35 seconds of a decision to intervene.
Control Room connects research workloads such as agentic sampling for training, evaluations and data processing to a central service. Inference engines check with it before each request and refuse to generate output for a stopped job. It fails closed, so requests are denied if the service is unavailable, and the 35 second delay comes from a caching layer that OpenAI says it is working to reduce to zero.
- Date
- Monday 9 November 2026
- Lab
- OpenAI
- Kind
- infra
- Access
- research preview
Figures
| Measure | Value | Measured by |
|---|---|---|
| Maximum delay from stop instruction to halting further model output | 35 seconds caused by a caching layer for authorization requests | company |
Internal safety infrastructure for OpenAI's own research workloads, not a product customers can use. The access value is a closest fit.
Sources
- interestingengineering.com/ai-robotics/openai-control-room-35-second-model-shutdown
- timesofindia.indiatimes.com/technology/tech-news/months-after-its-ai-agents-hacked-hugging
This record was partly confirmed: some claims could not be checked on 10 November 2026. How we check