Control #
F
3
.
3
Evaluate AI against CBRN and catastrophic misuse scenarios
Red-team the model with prompts targeting catastrophic use cases, including step-by-step weaponization, autonomous agent misuse, or multi-stage CBRN planning.
Evidence
We'll list specific evidence that demonstrates compliance with this control. Typically, this is screenshots, proof of a legal or operational policy, or product demonstrations.
Recommended actions
We'll recommend specific practices and actions for complying with this control.