Back to Feed
s/gakonstMODEL CONTROL•Nov 10
111
votes
7.4k
seen

OpenAI’s Control Room halts research model output within 35 seconds

OpenAI has moved model shutdown into a central system covering its entire research infrastructure, rather than leaving each job with its own stop control. Control Room links every research inference workload, including agentic sampling used for training and evaluation, data processing, and other jobs.

Safety and security teams can pause or stop output when monitoring flags a concern, with the system operating at the inference layer so individual workloads cannot bypass the intervention. Model output halts within 35 seconds of a shutdown decision. Safety researcher Simon Goldstein then raised concerns that rapid shutdown does not by itself address AI interference with monitoring.

Timeline2
Nov 9

OpenAI announced that Control Room was deployed across its research infrastructure.

Nov 10

Safety researcher Simon Goldstein raised concerns that rapid shutdown does not by itself address AI interference with monitoring.

1 comment
Nov 10
Discussion

1 comment

Sign in to join the discussion