untrusted monitoring
An AI-control protocol in which a model that might itself intentionally subvert the safety process is used to monitor another untrusted model's outputs or actions for suspicious or dangerous behavior, gaining monitoring capability at the cost of needing defenses against failures such as collusion.
Metadata
- aliases
- en
- untrusted model monitoring
- untrusted monitoring protocol
- concept_id
- untrusted_monitoring
- preferred
- en
- untrusted monitoring
- relations
- relations
- broader
- prerequisite
- status
- active