attack policy
In a control evaluation, a red-team-constructed replacement for an untrusted model that is elicited, prompted, scaffolded, or trained to carry out a strategy for causing a specified unacceptable outcome while avoiding detection or shutdown, so that the protocol can be tested against adversarial model behavior.
Metadata
- aliases
- en
- attack policies
- red-team attack policy
- concept_id
- attack_policy
- preferred
- en
- attack policy
- relations
- relations
- prerequisite
- related
- status
- active