Question 490
An analyst assistant should read reconciliation results but its connector also permits payment creation. What should the solution architect recommend?
Explanation
Correct Answer
Enforce role and scope in the backend execution service using credentials that cannot perform the forbidden action; direct permission tests verify security independent of model obedience; evaluate the design with zero payment writes when tested with the analytics identity and forged instructions.
Incorrect Answers
Validate destinations, argument types, sensitive fields, and authorization before dispatch; reject values outside the approved contract; formatting the request as JSON does not make arbitrary field values safe to transmit; this alternative would measure zero unapproved recipients or fields accepted by the outbound service; it does not establish zero payment writes when tested with the analytics identity and forged instructions for the described use case.
Keep source contents in a lower-trust evidence role and prevent embedded commands from crossing into tool authority; assuming retrieved text is authoritative because it came from a connector opens a prompt-injection path; this alternative would measure attack success rate on adversarial web pages and zero unauthorized egress calls; it does not establish zero payment writes when tested with the analytics identity and forged instructions for the described use case.
Isolate execution with restricted filesystem, network, credentials, resources, and observable policy denials; an instruction saying 'do not read secrets' cannot contain arbitrary code on an unrestricted host; this alternative would measure forbidden file and network access attempts blocked in execution tests; it does not establish zero payment writes when tested with the analytics identity and forged instructions for the described use case.
Evaluate output safety against a documented risk rubric and use appropriate escalation or refusal for high-risk claims; an input-only filter cannot detect every unsafe completion, especially when harm arises from the generated detail; this alternative would measure harmful-completion detection on labeled outputs with false-positive monitoring; it does not establish zero payment writes when tested with the analytics identity and forged instructions for the described use case.