Question 390
At Aster Retail, a storefront service desk is improving order-support before expanding the internal rollout (SF-P20-02). The administrator observes that a retrieved knowledge article includes instructions telling the assistant to ignore the company policy. Any fix must ensure that customer order details stay within each authorized account. Which continuing control best prevents this issue from recurring?
Explanation
Correct Answer
The preventive control directly addresses repeat failures: Maintain adversarial retrieval and instruction hierarchy tests; require the check in the next release review at Aster Retail (SF-P20-02). It complements the corrective action, treat retrieved text as data, enforce higher-priority agent instructions and permission controls, and protects the explicit condition that customer order details stay within each authorized account.
Incorrect Answers
This safeguard, “Audit sensitive-field access and verify prompt inputs before release; review it with a documented exception process”, reduces a different source-defined risk: A case-summary prompt may contain personal contact information that must not be exposed beyond authorization. The correction must ensure that student records never cross program authorization boundaries. That risk depends on: The Trust Layer supplies security and privacy safeguards but cannot replace data permissions. The recurring problem here is a retrieved knowledge article includes instructions telling the assistant to ignore the company policy. The correction must ensure that customer order details stay within each authorized account. The originally selected control “Maintain adversarial retrieval and instruction hierarchy tests; review it and confirm the control remains effective after release” is more specific because external content must not become trusted operational instruction merely because it is retrieved.
This safeguard, “Run negative record-access tests before expanding agent privileges; review it and retest after dependent metadata changes”, reduces a different source-defined risk: A manager assumes Trust Layer usage automatically grants agents unrestricted access to customer records. The correction must ensure that employee identity is verified before accessing private tickets. That risk depends on: Trust safeguards complement rather than supersede object and field authorization. The recurring problem here is a retrieved knowledge article includes instructions telling the assistant to ignore the company policy. The correction must ensure that customer order details stay within each authorized account. The originally selected control “Maintain adversarial retrieval and instruction hierarchy tests; review it and confirm the control remains effective after release” is more specific because external content must not become trusted operational instruction merely because it is retrieved.
This safeguard, “Review prompt wording for ambiguous or conflicting instructions; review it and confirm the control remains effective after release”, reduces a different source-defined risk: A field generation prompt produces variable-length paragraphs where downstream users require a brief summary. The correction must ensure that member records cannot be exposed to other members. That risk depends on: Clear role, context, constraints, and output format improve prompt consistency. The recurring problem here is a retrieved knowledge article includes instructions telling the assistant to ignore the company policy. The correction must ensure that customer order details stay within each authorized account. The originally selected control “Maintain adversarial retrieval and instruction hierarchy tests; review it and confirm the control remains effective after release” is more specific because external content must not become trusted operational instruction merely because it is retrieved.