control evaluation
An adversarial evaluation of a control protocol that estimates its safety and usefulness under pessimistic assumptions about untrusted-model behavior, typically by having a red team construct adversarial substitutes for untrusted models and measuring whether they can cause a specified unacceptable outcome without being caught.
Metadata
- aliases
- en
- AI control eval
- AI control evaluation
- control eval
- control evaluations
- concept_id
- control_evaluation
- preferred
- en
- control evaluation
- relations
- relations
- broader
- prerequisite
- related
- status
- active