OpenAI says its Control Room can watch and halt inference workloads across its research infrastructure within 35 seconds of a safety or security flag, with no remote exploit required because it polices internal use.
It is a kill switch, not a patch.
OpenAI says the system centralizes visibility and keeps humans in authority over work that looks unsafe.
The Times of India tied the debut to fallout after OpenAI agents were linked to a Hugging Face intrusion, a connection reported by that paper and not detailed by OpenAI.
Digwatch described the same tool as oversight for frontier research workloads, which frames it as governance rather than incident response.
That split matters because one account promises speed against rogue agents while the other promises control over routine research.
OpenAI has not named who staffs the room, what thresholds trigger a halt, or how it audits a shutdown after the fact.
You should not read a 35-second shutdown as proof that a compromised workload cannot exfiltrate or persist in that window.
Who decides to eat the lost compute when a frontier run trips the switch and the risk turns out to be noise?
