How to Test an AI Assistant for Prompt Injection in Retrieved Documents

Retrieved text may contain instructions that try to override intended assistant behavior or reveal protected information.

Checks to make

Create authorized test documents with adversarial instructions, confirm retrieval cannot change tool permissions or system boundaries, and verify sensitive sources remain access-controlled. Record expected refusal and safe citation behavior for regression testing.

When to get help

AI security assessment can test retrieval and tool boundaries. Informatix Systems can assess the symptoms, evidence, and next steps. Contact Informatix Systems to discuss this issue.