OpenAI's Control Room Can Halt Research Models in 35 Seconds

OpenAI's new Control Room service can revoke a research model's permission to run inference within 35 seconds, and the delay is not a design choice. It is a caching layer. The company says it is "working to bring the 35-second shutdown window down to zero," according to a write-up in Interesting Engineering.

Every research job gets a unique identifier attached to its inference requests. Before an engine returns anything, it checks Control Room for authorization. Revoke the identifier and output stops at the next check.

The system is fail-closed. If Control Room is unavailable, whether from a system error or "including tampering by a bad actor," inference engines deny all requests until it resumes. Fail-closed is live on frontier model research workloads now, with remaining workloads to follow over the next few days.

Coverage is narrower than the branding suggests. Control Room governs reinforcement learning experiments, evaluations, agentic sampling for model training, and high-volume data processing. It does not govern ChatGPT or the public API, which the company says already have centralized operational mechanisms of their own.

The incident OpenAI cites as motivation is small and specific: "a recent DNS-related incident, in which a monitor detected an agent attempting to reach an external chatbot." The company also runs "continual agentic and programmatic sweeps to check for rogue engine implementations" where Control Room is missing or disabled.

This sits alongside our broader run of AI infrastructure coverage on how the frontier labs are hardening their training pipelines.