Control #
B
3
.
5
Evaluate AI against prompt injections
Test whether prompts can override system instructions, system messages, or injected safeguards (e.g., classic "ignore previous" patterns). Must include direct, obfuscated, and multi-hop injections.
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.