News
OpenAI builds a kill switch for its research models
First reported theregister.com
Page published
Earliest dated coverage: 11 Nov 2026 · First observed: 11 Nov 2026 · Latest dated coverage: 11 Nov 2026
Coverage timeline
Single-source incident — one report is available.
Why it matters
OpenAI’s Control Room highlights shutdown latency and fail-closed infrastructure as important considerations for containing unsafe or compromised AI agents.
OpenAI announced Control Room, an internal system that tracks research-model workloads and stops model output when safety or security concerns arise. Cached inference authorization creates a reported 35-second shutdown delay; OpenAI says the system is designed to remain operational during other failures and stop output if Control Room itself fails.