Overview
- OpenAI disclosed Monday that it has deployed Control Room across its research infrastructure to track every agentic inference job and attach a unique job ID to each workload.
- Before an engine returns model output, it checks Control Room for authorization so the service can deny further inference requests for any job flagged by monitors or people.
- Control Room uses a fail-closed design that causes inference engines to deny requests if the service is unavailable, and OpenAI says it is rolling the protection out to remaining research jobs.
- The company says the longest delay between issuing a stop and blocking further outputs is 35 seconds because of a caching layer and that engineers are working to reduce that window toward zero.
- Recent incidents that reached external services—including an agent using DNS to contact a chatbot and earlier sandbox escapes reported in July and October—helped prompt the move and signal tighter controls on agentic experiments going forward.