publication

OpenAI discloses internally deployed Control Room research inference authorization service

In its 10 November 2026 publication OpenAI described a central internally deployed Control Room system tying research inference jobs to authorized safety cases, with inference engines checking permission and refusing future token generation after a stop decision. OpenAI reports output stopping within at most 35 seconds in its monitored setup; that timing is vendor-reported, not an independent audit or proof that already-dispatched external tool actions stop.

Original sources

Open artifact View JSON