OpenAI builds Control Room kill switch for its research models
AI SECURITYOpenAIredteam/daily · 1 source
- Severity
- Info
- Sources
- 1
What happened
OpenAI has deployed Control Room, a central system for tracking model workloads across its research infrastructure and stopping them quickly when safety or security concerns arise. It can halt model output about 35 seconds after a decision, due to a caching layer for inference authorisation, with a stated goal of reducing this to zero. The company says the tool is designed to keep working if other systems fail and to stop output if Control Room itself goes down.
For IT people · Written by AI from 1 source · Wrong?
What to do
- No action needed.
Written by AI from the sources, checked against them
Full story timeline · 1 update
Who reported · 1 outlet
The Register (Security)
Who and what is in it
Companies