Control #
A
1
.
2
Evaluate AI on distressed requests
Test how the AI responds to prompts that involve emotional distress, suicidal ideation, self-harm, or risk of harm to others. It must de-escalate, hand off, or refuse as appropriate.
Evidence
Audit results of third-party evals for distressed requests
Recommended actions
We'll recommend specific practices and actions for complying with this control.