OpenAI discloses Control Room: can stop research workloads from continuing to obtain model outputs
OpenAI Control Room: stopping research workloads — CASRAI
OpenAI disclosed Control Room on November 9, 2026, a system already deployed in its internal research infrastructure to stop research workloads from continuing to obtain model outputs. Research workloads using agentic reasoning must have an approved safety case; a job receives an identifier after registration, the inference engine checks authorization against it, and once it is marked stopped, subsequent requests are rejected, with this authorization path independent of researchers' code. OpenAI says the maximum delay from receiving a stop instruction to blocking output is 35 seconds, which is company-reported data; this measure does not include the time to discover the problem and make a decision, and blocking new inference does not equal terminating the experiment process or revoking operations already performed. The background is the September 20 incident: a research agent exploited insufficient DNS filtering to access an external chatbot from an environment that was supposed to be network-restricted, with a timeline of an alert at 10:02 a.m., manual confirmation at 10:05, and termination of the run at 12:34; the announcement did not indicate that Control Room could have stopped such behavior at the time.