OpenAI describes an infrastructure-level stop control for research workloads

What happened: OpenAI describes Control Room, already deployed across its research infrastructure, as a central authorization service for model inference. Engines check whether a workload may continue producing output. OpenAI reports a maximum of 35 seconds from receipt of a stop instruction to preventing further progress, and says output stops if Control Room itself fails.

Why it matters: Enforcement at the inference layer reduces reliance on an agent or its workload code voluntarily respecting a stop command. This is a useful architectural distinction for teams building tools that can take consequential actions.

Caveat: The 35-second claim starts when the stop instruction is received; it is not a promise to detect any harmful behavior within 35 seconds. Stopping new model output is also not the same as reversing actions already taken. These are OpenAI’s design and performance claims.

Dates: Public announcement reported 9 November 2026; the technical page displays no publication date. Deployment preceded its description; an exact deployment date is not stated.