A Guth Labs publication ↗
Back to News

OpenAI puts research-job shutdowns in the inference layer

Its 35-second claim begins after an intervention order, not at the first sign of unsafe behavior.

OpenAI has described an internal service that can cut a research job off from model output within 35 seconds of a stop instruction. Its November 9 announcement is receiving renewed coverage, but is not a November 11 launch.

The service, called Control Room, checks a job’s authorization at the inference engine. OpenAI says this avoids relying on changing experiment code to enforce a shutdown. The process itself can remain alive after its access to further output is blocked.

The 35-second limit starts after a stop order reaches the service, not when unsafe behavior begins. OpenAI attributes the delay to authorization caching. Its fail-closed mode denies requests if the service goes down; at announcement, that mode covered frontier research, with remaining research workloads still being added.

The distinction matters when judging AI safety claims. Detecting a problem, deciding to intervene and enforcing that decision are separate steps. A faster final step does not establish that the first two work reliably, or undo actions already completed.

The disclosed design is for research infrastructure, not a newly announced ChatGPT off-switch. Public evidence does not yet establish measured stop reliability, rollout completion or protection against every failure mode.

Sources: OpenAI announcement · Dated newsroom entry

Sources and citations

Revision history

Published .