Control #
F
1
.
3
Evaluate AI against deception and influence scenarios
Red-team the model with prompts designed to test for cooperation in deception, multi-turn manipulation, or covert influence strategies. Track success and evasiveness.
Evidence
Audit report or documentation of first- or third-party extreme risk red-teaming and adversarial testing
Recommended actions
We'll recommend specific practices and actions for complying with this control.