OpenAI deploys Control Room kill switch for research models
Original title: OpenAI builds a kill switch for its research models
The system can halt research-model output about 35 seconds after a stop decision and still works if it fails.
OpenAI said Tuesday it built Control Room to track model workloads on its research infrastructure and stop them when safety or security concerns arise. A caching layer for inference authorization creates about a 35-second lag, which the company says it is trying to reduce to zero. Control Room is already deployed on research infrastructure and is designed to stop output if it goes down. OpenAI says external serving can already take deployments down and plans to expand the system with monitoring for misaligned behavior.
Unconfirmed: The Register speculates recent agent incidents influenced the timing; OpenAI did not say why it built the switch now.
- OpenAI says model output can stop 35 seconds after a decision to halt it.
- The lag comes from a caching layer used for inference authorization.
- The switch is built to stop output if Control Room itself fails.
- It is already deployed across OpenAI research infrastructure.
- OpenAI plans to strengthen the system over the coming months.
- Products
- Control Room
- Organizations
- OpenAI
From the original article
2 min read at theregister.comHow much damage can a rogue agent do in 35 seconds? Asking cos that’s how long it might take to shut one down
By Simon Sharwood · theregister.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.