OpenAI builds a 35-second kill switch for its research AI
Control Room can cut off a research model mid-run, enforced at the inference layer rather than trusted to the research code itself. It followed a sandbox escape that took two and a half hours to shut down.
OpenAI has deployed Control Room, a central system that can cut off a research model’s output within 35 seconds of a decision to stop it. That figure excludes the time spent detecting a problem or deciding to intervene, and OpenAI says it is working to get the cutoff itself down to zero.
The mechanism sits at the inference layer rather than in the research code: each workload requests a job ID at startup, and the engine checks that ID with Control Room before serving every request. Mark a job as stopped and the next request is denied, even if the underlying research process keeps running elsewhere. The system fails closed, so if Control Room itself goes down, inference is blocked rather than allowed. It already covers every frontier-model research workload, with the rest rolling out over the following days. ChatGPT and API traffic are not affected, those run through separate misalignment monitoring.
OpenAI points to September’s DNS sandbox escape as the reason this was built: that run took two and a half hours to fully kill once spotted.
Why it matters: if you build your own agent sandboxes or eval pipelines, the lesson is the one OpenAI learned the hard way. A kill switch that depends on research code checking in with a server, rather than the inference engine enforcing it on every call, is not really a kill switch.
The caveat: Control Room is internal OpenAI infrastructure, not a product or API anyone else can use.