OpenAI Control Room adds 35-second kill switch

This is less a dashboard than an infrastructure-level circuit breaker for increasingly autonomous research workloads.

• –Putting termination in the inference layer avoids relying on varied, potentially buggy workload code.
• –The 35-second maximum is a meaningful improvement, but OpenAI is still targeting instantaneous shutdowns.
• –Fail-closed behavior improves containment while creating a real availability tradeoff: a control-plane outage can pause every protected workload.
• –The design directly addresses prior operational gaps where detection occurred long before a run was actually stopped.