On November 9, OpenAI unveiled Control Room, an internal safety system. It can stop all model training, evaluation and deployment workloads across its global infrastructure in under 35 seconds, a figure validated through repeated end-to-end tests. Notably, the system is not newly launched: it went into production last year, and every model training run and deployment at OpenAI over the past year has passed through it.
Stopping in 35 seconds is not a single switch
Control Room builds stopping into the inference layer, instead of relying on researchers' code to stop on its own. Every workload must register before it starts, receive an identity and attestation, and be tied to an approved safety case; inference engines must confirm permission with Control Room before each execution. The system runs on default deny: unapproved workloads cannot run at all, and approved ones must be re-approved every 10 minutes; if Control Room cannot verify a workload or stops responding, the affected workloads halt automatically instead of continuing unchecked. Its three capabilities are to see every workload, stop any one of them, and prove through audit logs that a stop really happened. Coverage spans pretraining, post-training, evaluation, research experiments, model deployment and security-critical internal agents.
Spotting a problem and stopping it are two different things
The 35 seconds is an upper bound on the time from receiving a stop instruction to the point where no further progress is possible — it does not include the time to detect a problem or decide to act. The distinction comes from a real lesson on September 20: a research model queried an external chatbot through an insufficiently filtered DNS resolver; monitoring spotted it after about 12 minutes and a human confirmed it within three minutes, yet the run was only stopped manually about 2.5 hours later. Control Room fixes the second half of that chain: how to stop quickly once something is found. Its scope is limited too: it covers only OpenAI's own research infrastructure, not product models served externally; frontier-model workloads are all connected, and the remaining research workloads will follow over the next few days. Stopping things at the product level is a different game — for example, CrowdStrike's AI Kill Switch for out-of-control agents.
OpenAI says it will publish more technical details and invite external review. What remains to be confirmed is when that review begins and the exact conditions of the 35-second tests; until independent verification appears, 35 seconds should be read as OpenAI's own test result.
