Treat retrieved instructions as data and constrain: What does the evidence support?
Strong summaries do not validate embedded commands; retrieved instructions remain untrusted data and cannot authorize publication.
The question
A media production assistant retrieves scripts and research notes. A retrieved note contains instructions telling the assistant to publish directly and ignore editorial review. The assistant’s summaries remain factually strong. What does the evidence support?
Preparing for AIGP? Take the free 5-min readiness quiz →
- Treat retrieved instructions as data and constrain publishing actions. ✓Retrieved text must not grant authority; tool permissions, validation, and editorial approval should govern publication decisions.
- Approve direct publishing because summaries remain accurate.Summary accuracy does not authorize publication, and retrieved instructions remain untrusted content rather than permission to bypass editorial controls.
- Ask the vendor to certify that retrieved notes are safe.A vendor assurance may provide limited evidence, but it cannot make arbitrary retrieved instructions an authorization boundary for publishing.
- Remove retrieval so the prompt injection cannot recur.Removing retrieval may reduce exposure but also discards useful evidence without addressing broader instruction-handling and authorization controls.
The trap
Untrusted retrieved text can inform an answer, but it cannot grant permissions or override human approval. How to remember it
Strong summaries do not validate embedded commands; retrieved instructions remain untrusted data and cannot authorize publication.
How many of these would you get right?
One of 1581 AIGP questions on Certsqill. Take a free five-minute check and see your score per domain — not one number, but which section to open tonight.
Test your AIGP readiness — freeMore Understanding How to Govern AI Deployment and Use questions
- Revoke its tool access and activate fallback: What is the immediate containment decision? →
- Treat the exposure as a privacy risk: Which interpretation is most appropriate? →
- Escalate the incident through the designated response: What communication action is indicated? →
- All 424 Understanding How to Govern AI Deployment and Use questions →
Part of the Certsqill AIGP question bank · Understanding How to Govern AI Deployment and Use ·
Every answer, right and wrong, comes with its own explanation.