OpenAI describes Control Room, a system to stop research model jobs within 35 seconds

OpenAI · Company post

Read the source (opens in a new tab)
What happened
OpenAI published a post on November 9, 2026 describing Control Room, a central system that tracks model workloads across its research infrastructure and can stop them.
Key points
  • Every research job gets an ID, and inference engines check with Control Room before each request; a stopped job's requests are refused, without relying on researchers' code.
  • OpenAI says it halts model output "within 35 seconds of a decision to intervene," a lag from a caching layer it is working to bring to zero.
  • Its misalignment monitors, which run on all frontier reinforcement learning jobs, can stop a job through Control Room; OpenAI says one caught its recent DNS-related incident.
  • The system "fails closed": if Control Room goes down, including through tampering, all inference requests are denied; this covers frontier research now and the rest within days.
Evidence
Company description of its own internal safety infrastructure. It reports no count of jobs stopped and no outside testing of the system.

See it in the latest record