Practice question 489 of 498 · hard
An approved AI agent retrieves support documents and can issue customer refunds. Prompt Shields detects adversarial text, but the refund tool accepts any customer account and amount presented by the agent. Which architecture best satisfies least privilege and human oversight?
Choose the single best answer.
Show the answer and explanation
✓ Correct answer
- Authorize customer and refund scope in the tool service and enforce approval before designated high-impact refunds
Authorize customer and refund scope in the tool service and enforce approval before designated high-impact refunds: Independent constraints limit execution even if a model follows malicious retrieved instructions. Repeat the refund limit in the safety prompt without changing the tool: Natural-language instructions leave the underlying excessive authority intact. Approve every tool request automatically if Prompt Shields returned no detection: Detection does not establish transaction authorization or complete safety. Give the agent more data rights so it can independently decide all exceptions: Broader authority increases the consequences of agent manipulation. Review only a daily refund report after irreversible actions finish: Retrospective review does not provide the required pre-action approval boundary.
Read and review here, or open the app to practise and save your results on this device.
Unofficial study material. Not affiliated with or endorsed by Microsoft. These are original learning questions, not actual exam questions. Product capabilities and exam objectives can change; verify current Microsoft documentation. Practice scores do not predict or guarantee an exam result.