Prompt Injection in Support AI
Question Variations
- "Why can't prompt wording alone prevent injection?"
- "How would you test an agent against retrieved malicious text?"
- "How would you secure a support assistant that reads untrusted emails before using account tools?"
Why This Is Asked
This tests whether a candidate understands prompt injection as an application-security problem rather than a prompt-writing problem. Interviewers assess untrusted-context handling, capability controls, authorization, and adversarial evaluation.
Key Concepts
- Untrusted context: Retrieved text is data, not privileged instruction.
- Capability controls: Sensitive actions require server-side authorization and confirmation.
- Data minimization: Give the model only the context and tools required.
- Testing: Red-team direct and indirect injections continuously.
Question Variations
- “Why can’t prompt wording alone prevent injection?”
- “How would you test an agent against retrieved malicious text?”
- “How would you secure a support assistant that reads untrusted emails before using account tools?”