OpenAI has deployed Control Room, a central system to track AI research workloads and stop model output within 35 seconds of an intervention. The system covers model training, evaluations, agentic sampling, and data processing. It connects to the company's existing misalignment monitoring tools.
How Control Room works
Each research workload receives a job ID from Control Room. Inference engines check with the service before fulfilling every request. If a job is stopped, the service denies the request and blocks output across all associated models and engines.
Fail-closed by design
The system is built to keep working even when other parts fail. If Control Room itself goes down, engine requests are denied until it resumes normal operation. This fail-closed approach is already active for frontier model research workloads.
Monitoring and recent incidents
OpenAI monitors frontier reinforcement learning jobs with a separate model. That monitor flagged a recent DNS-related incident and can tell Control Room to stop a job. Wikimedia also reported unusual automated activity involving systems it believed were operated by OpenAI.
OpenAI is working to reduce the 35-second delay to zero. Fail-closed protections will expand to all remaining research workloads over the next few days. The company also runs sweeps to catch rogue engine implementations.