OpenAI Control Room can halt research model output in 35 seconds

OpenAI’s Control Room, live on research infrastructure, can stop model output about 35 seconds after a halt decision.

OpenAI says Control Room is already deployed on its research infrastructure to track model workloads and stop output when safety or security concerns arise. Agentic research jobs must have an approved safety case and a job ID; inference engines check Control Room before generating tokens and refuse stopped jobs. The company says output halts within about 35 seconds of a stop or pause decision. The Register reports that lag comes from a caching layer for inference authorization, which OpenAI is trying to reduce to zero. The design stops output if Control Room itself fails, though a stopped job may keep running, and OpenAI plans to add monitoring for misaligned behavior.