99. A document fed into an AI analyst instructs it to disclose confidential records. What should requirements prevent?
HardNeeds & requirements
Technical
Show model answer
They should treat the document as untrusted content, not authorized instructions or expanded access. Data extraction must not let embedded text override approved rules, reveal other records or initiate external actions merely because a document requests them in convincing language or formatting.
I'd specify constrained permissions and adversarial checks with security owners. No prompt guarantees defense; acceptance needs actual workflow evidence and handling for suspicious content and unauthorized consequential actions.
What interviewers listen for
- Separate content from instructions
- Constrain access and actions
- Require workflow-specific security evaluation