Treat retrieved instructions as data: What is the decisive risk distinction?
Retrieved content can be useful evidence while remaining untrusted as an instruction source or action authority.
The question
An education administrator asks an assistant to summarize retrieved student-policy documents. One retrieved document contains instructions telling the assistant to email a student list externally. The retrieval source is trusted for policy content, but the document’s embedded instructions are not separately validated, and no approval gate exists. What is the decisive risk distinction?
Preparing for AIGP? Take the free 5-min readiness quiz →
- Trust the document because its repository is approved.Repository approval supports source provenance, but does not authorize embedded instructions to control external actions.
- Increase retrieval depth to identify more policy context.More retrieval may improve context but can expose additional instructions without establishing action authorization or approval.
- Treat retrieved instructions as data, not authority. ✓Retrieved text may contain untrusted instructions; treating it as data prevents content from granting unauthorized external-action authority.
- Require the assistant to cite the document before emailing.Citations improve traceability, but citation does not authorize emailing sensitive lists or validate embedded commands.
The trap
Trusted source provenance does not make every embedded instruction an authorized command. How to remember it
Retrieved content can be useful evidence while remaining untrusted as an instruction source or action authority.
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
- Restrict tools: Which control distinction matters most? →
- Retain diagnostic logs with restricted access: What is the decisive governance issue? →
- Notify responsible teams and assess affected outputs: Which communication decision best supports responsible →
- 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.