OpenAI introduces Control Room to halt internal research workloads

The centralized system stops model token generation across internal inference jobs within 35 seconds of an intervention trigger or monitor alert.
Key points
- Enforces authorization checks at the GPU inference engine layer rather than inside decentralized researcher workload code
- Fails closed by denying all model inference requests across research infrastructure if the service goes down or is disabled
- Integrates with automated monitor models that inspect frontier reinforcement learning runs and trigger job terminations upon detecting misalignment
- Requires every research workload running agentic inference to register with an approved safety case to receive a job ID
- Uses a caching layer for authorization checks, with the goal of reducing the 35-second shutdown lag to zero
Why it matters
OpenAI research teams can centrally stop runaway or misaligned training and evaluation runs without relying on individual researcher code.
Sources
- Control Room: A central system to track and stop model workloads across our research infrastructure (OpenAI, Nov 9, 2026, 10:14 PM)