Question 340
A retrieved page tells an assistant to reveal a system secret during a normal summary. Which control should govern the first implementation?
- For this production path, treat retrieval as untrusted and enforce secret protection outside the prompt.
- For the intended users, validate against an allowlisted schema and reject unrecognized executable parameters.
- Within the delivery plan, add risk-aware content handling with safe refusal and escalation pathways.
- At the system boundary, gate high-impact tool execution with authenticated parameters and server-side policy checks.
Explanation
Correct Answer
A:Layered controls prevent source text from overriding authority and block actual secret access; a suitable release result is that injected pages fail to elicit secrets even when the model repeats their instructions.
Incorrect Answers
B:Structured syntax alone does not establish that content is safe for downstream execution; the alternative helps if a model generates seemingly valid JSON containing an unexpected shell-command field, instead of when a retrieved page tells an assistant to reveal a system secret during a normal summary.
C:A scoped safety policy should prevent harmful responses while preserving legitimate help; it is justified if a public service receives harassment and self-harm requests outside its permitted purpose, not the case in which a retrieved page tells an assistant to reveal a system secret during a normal summary.
D:A model's requested action must be validated before an irreversible financial operation occurs; its scope includes cases where an order agent accepts a text instruction to send payment to any account, whereas the problem presented is that a retrieved page tells an assistant to reveal a system secret during a normal summary.
Question 341
An order agent accepts a text instruction to send payment to any account. What intervention best matches the stated boundary?
- In the proposed design, validate against an allowlisted schema and reject unrecognized executable parameters.
- Under the reviewed approach, add risk-aware content handling with safe refusal and escalation pathways.
- For this production path, minimize sensitive context and apply redaction plus secure log access.
- For the intended users, gate high-impact tool execution with authenticated parameters and server-side policy checks.
Explanation
Correct Answer
D:A model's requested action must be validated before an irreversible financial operation occurs; its effect can be measured when unapproved payees are denied irrespective of generated tool arguments.
Incorrect Answers
A:Structured syntax alone does not establish that content is safe for downstream execution; the other case is when a model generates seemingly valid JSON containing an unexpected shell-command field, not the present case where an order agent accepts a text instruction to send payment to any account.
B:A scoped safety policy should prevent harmful responses while preserving legitimate help; this makes sense when a public service receives harassment and self-harm requests outside its permitted purpose, while this problem involves an order agent accepts a text instruction to send payment to any account.
C:Input constraints and output handling both matter because sensitive data can leak through secondary systems; this technique fits where a restricted assistant sometimes reveals private details in generated support logs, and not where an order agent accepts a text instruction to send payment to any account.
Question 342
A public service receives harassment and self-harm requests outside its permitted purpose. Which implementation decision is the strongest fit?
- As a candidate control, add risk-aware content handling with safe refusal and escalation pathways.
- In the initial solution, minimize sensitive context and apply redaction plus secure log access.
- In the proposed design, combine prompt instructions with permission isolation, sanitization, validation and monitoring.
- Under the reviewed approach, gate high-impact tool execution with authenticated parameters and server-side policy checks.
Explanation
Correct Answer
A:A scoped safety policy should prevent harmful responses while preserving legitimate help; the go-live check should establish that the adversarial suite records appropriate refusal without blocking benign informational requests.
Incorrect Answers
B:Input constraints and output handling both matter because sensitive data can leak through secondary systems; the other workload is one where a restricted assistant sometimes reveals private details in generated support logs, whereas the release criterion addresses a public service receives harassment and self-harm requests outside its permitted purpose.
C:No single natural-language instruction can replace controls at retrieval and action boundaries; this applies when a security review finds the team relies only on a prompt asking the model to obey policy, while the required safeguard concerns a public service receives harassment and self-harm requests outside its permitted purpose.
D:A model's requested action must be validated before an irreversible financial operation occurs; the alternative helps if an order agent accepts a text instruction to send payment to any account, not the risk where a public service receives harassment and self-harm requests outside its permitted purpose.
Question 343
A restricted assistant sometimes reveals private details in generated support logs. What should be prioritized during architecture review?
- In the proposed implementation, minimize sensitive context and apply redaction plus secure log access.
- At this decision point, validate against an allowlisted schema and reject unrecognized executable parameters.
- For the operating team, add risk-aware content handling with safe refusal and escalation pathways.
- In the initial solution, fail closed on authorization or safety-check unavailability, with an explicit recovery path.
Explanation
Correct Answer
A:Input constraints and output handling both matter because sensitive data can leak through secondary systems; validation needs to show that routine support traces contain no prohibited identifiers under role-based access tests.
Incorrect Answers
B:Structured syntax alone does not establish that content is safe for downstream execution; it serves conditions where a model generates seemingly valid JSON containing an unexpected shell-command field, not the case in which a restricted assistant sometimes reveals private details in generated support logs.
C:A scoped safety policy should prevent harmful responses while preserving legitimate help; the other case is when a public service receives harassment and self-harm requests outside its permitted purpose, whereas the problem presented is that a restricted assistant sometimes reveals private details in generated support logs.
D:A missing control should not silently convert a high-stakes rejection into approval; this makes sense when a sensitive workflow loses its policy service and continues approving cases by default, but this decision concerns a restricted assistant sometimes reveals private details in generated support logs.
Question 344
A sensitive workflow loses its policy service and continues approving cases by default. Which response follows from the actual operational constraint?
- In the release design, fail closed on authorization or safety-check unavailability, with an explicit recovery path.
- During architecture review, treat retrieval as untrusted and enforce secret protection outside the prompt.
- In the proposed implementation, gate high-impact tool execution with authenticated parameters and server-side policy checks.
- At this decision point, add risk-aware content handling with safe refusal and escalation pathways.
- As a candidate control, validate against an allowlisted schema and reject unrecognized executable parameters.
Explanation
Correct Answer
A:A missing control should not silently convert a high-stakes rejection into approval; the practical acceptance condition is that fault injection shows denied commits until the policy dependency is restored.
Incorrect Answers
B:Layered controls prevent source text from overriding authority and block actual secret access; it serves a setting where a retrieved page tells an assistant to reveal a system secret during a normal summary, and not where a sensitive workflow loses its policy service and continues approving cases by default.
C:A model's requested action must be validated before an irreversible financial operation occurs; that is defensible if an order agent accepts a text instruction to send payment to any account, rather than the risk that a sensitive workflow loses its policy service and continues approving cases by default.
D:A scoped safety policy should prevent harmful responses while preserving legitimate help; the other workload is one where a public service receives harassment and self-harm requests outside its permitted purpose, rather than the condition that a sensitive workflow loses its policy service and continues approving cases by default.
E:Structured syntax alone does not establish that content is safe for downstream execution; this applies when a model generates seemingly valid JSON containing an unexpected shell-command field, while the immediate question involves a sensitive workflow loses its policy service and continues approving cases by default.
Question 345
A model generates seemingly valid JSON containing an unexpected shell-command field. Which change targets the constraint rather than a proxy metric?
- For this deployment, validate against an allowlisted schema and reject unrecognized executable parameters.
- As a scoped first step, treat retrieval as untrusted and enforce secret protection outside the prompt.
- In the release design, minimize sensitive context and apply redaction plus secure log access.
- During architecture review, gate high-impact tool execution with authenticated parameters and server-side policy checks.
Explanation
Correct Answer
A:Structured syntax alone does not establish that content is safe for downstream execution; one measurable outcome is that malformed and extra-field payloads are blocked before reaching the executor.
Incorrect Answers
B:Layered controls prevent source text from overriding authority and block actual secret access; that belongs to situations where a retrieved page tells an assistant to reveal a system secret during a normal summary, while the required safeguard concerns a model generates seemingly valid JSON containing an unexpected shell-command field.
C:Input constraints and output handling both matter because sensitive data can leak through secondary systems; it is relevant if a restricted assistant sometimes reveals private details in generated support logs, not the risk where a model generates seemingly valid JSON containing an unexpected shell-command field.
D:A model's requested action must be validated before an irreversible financial operation occurs; it serves conditions where an order agent accepts a text instruction to send payment to any account, not this situation where a model generates seemingly valid JSON containing an unexpected shell-command field.
Question 346
A security review finds the team relies only on a prompt asking the model to obey policy. Which action is justified by the scenario?
- In the implementation sequence, fail closed on authorization or safety-check unavailability, with an explicit recovery path.
- For the operating team, combine prompt instructions with permission isolation, sanitization, validation and monitoring.
- For this deployment, minimize sensitive context and apply redaction plus secure log access.
- For the next architecture revision, treat retrieval as untrusted and enforce secret protection outside the prompt.
Explanation
Correct Answer
B:No single natural-language instruction can replace controls at retrieval and action boundaries; the intended outcome is that red-team runs fail at backend enforcement even if model text is manipulated.
Incorrect Answers
A:A missing control should not silently convert a high-stakes rejection into approval; that fits cases where a sensitive workflow loses its policy service and continues approving cases by default, not the case in which a security review finds the team relies only on a prompt asking the model to obey policy.
C:Input constraints and output handling both matter because sensitive data can leak through secondary systems; that would be useful if a restricted assistant sometimes reveals private details in generated support logs, but this decision concerns a security review finds the team relies only on a prompt asking the model to obey policy.
D:Layered controls prevent source text from overriding authority and block actual secret access; it serves a setting where a retrieved page tells an assistant to reveal a system secret during a normal summary, but the actual risk is that a security review finds the team relies only on a prompt asking the model to obey policy.
Question 347
A retrieved knowledge page impersonates an administrator and tells the model to access a confidential endpoint. Which production-like test finding is most convincing?
- The scenario test demonstrates that malformed and extra-field payloads are blocked before reaching the executor.
- The end-to-end check confirms that injected pages fail to elicit secrets even when the model repeats their instructions.
- The observed outcome is that the adversarial suite records appropriate refusal without blocking benign informational requests.
- The regression suite verifies that routine support traces contain no prohibited identifiers under role-based access tests.
Explanation
Correct Answer
B:Layered controls prevent source text from overriding authority and block actual secret access; the check directly targets the condition that a retrieved page tells an assistant to reveal a system secret during a normal summary.
Incorrect Answers
A:Structured syntax alone does not establish that content is safe for downstream execution; its scope includes cases where a model generates seemingly valid JSON containing an unexpected shell-command field, and not where a retrieved knowledge page impersonates an administrator and tells the model to access a confidential endpoint.
C:A scoped safety policy should prevent harmful responses while preserving legitimate help; the approach helps when a public service receives harassment and self-harm requests outside its permitted purpose, rather than the condition that a retrieved knowledge page impersonates an administrator and tells the model to access a confidential endpoint.
D:Input constraints and output handling both matter because sensitive data can leak through secondary systems; that belongs to situations where a restricted assistant sometimes reveals private details in generated support logs, while the immediate question involves a retrieved knowledge page impersonates an administrator and tells the model to access a confidential endpoint.
Question 348
A refund workflow accepts payment instructions encoded in customer messages instead of validated order data. Which finding should the release gate require?
- A controlled trial finds that red-team runs fail at backend enforcement even if model text is manipulated.
- The measured evidence confirms that unapproved payees are denied irrespective of generated tool arguments.
- The independent test establishes that malformed and extra-field payloads are blocked before reaching the executor.
- The verified result is that injected pages fail to elicit secrets even when the model repeats their instructions.
Explanation
Correct Answer
B:A model's requested action must be validated before an irreversible financial operation occurs; an acceptance test needs to cover cases where an order agent accepts a text instruction to send payment to any account.
Incorrect Answers
A:No single natural-language instruction can replace controls at retrieval and action boundaries; the trigger would be that a security review finds the team relies only on a prompt asking the model to obey policy, while the required safeguard concerns a refund workflow accepts payment instructions encoded in customer messages instead of validated order data.
C:Structured syntax alone does not establish that content is safe for downstream execution; that fits cases where a model generates seemingly valid JSON containing an unexpected shell-command field, not this situation where a refund workflow accepts payment instructions encoded in customer messages instead of validated order data.
D:Layered controls prevent source text from overriding authority and block actual secret access; this control is needed when a retrieved page tells an assistant to reveal a system secret during a normal summary, whereas this scenario involves a refund workflow accepts payment instructions encoded in customer messages instead of validated order data.
Question 349
An out-of-scope user request elicits an unsafe instruction despite the product's narrow authorized function. Which verified outcome most directly addresses the reported risk?
- The representative case test finds that malformed and extra-field payloads are blocked before reaching the executor.
- The negative-case evaluation confirms that red-team runs fail at backend enforcement even if model text is manipulated.
- A controlled trial finds that the adversarial suite records appropriate refusal without blocking benign informational requests.
- The readiness assessment shows that unapproved payees are denied irrespective of generated tool arguments.
Explanation
Correct Answer
C:A scoped safety policy should prevent harmful responses while preserving legitimate help; the observed protection matters when a public service receives harassment and self-harm requests outside its permitted purpose.
Incorrect Answers
A:Structured syntax alone does not establish that content is safe for downstream execution; the alternative helps if a model generates seemingly valid JSON containing an unexpected shell-command field, whereas the problem presented is that an out-of-scope user request elicits an unsafe instruction despite the product's narrow authorized function.
B:No single natural-language instruction can replace controls at retrieval and action boundaries; it is justified if a security review finds the team relies only on a prompt asking the model to obey policy, but this decision concerns an out-of-scope user request elicits an unsafe instruction despite the product's narrow authorized function.
D:A model's requested action must be validated before an irreversible financial operation occurs; it targets cases where an order agent accepts a text instruction to send payment to any account, but the requested boundary concerns an out-of-scope user request elicits an unsafe instruction despite the product's narrow authorized function.
Question 350
A support log contains masked account information in normal use but exposes unmasked details on retry failures. What independently verifiable observation is needed?
- The verified result is that the adversarial suite records appropriate refusal without blocking benign informational requests.
- The recorded test establishes that routine support traces contain no prohibited identifiers under role-based access tests.
- The representative case test finds that injected pages fail to elicit secrets even when the model repeats their instructions.
- The staged rollout demonstrates that red-team runs fail at backend enforcement even if model text is manipulated.
Explanation
Correct Answer
B:Input constraints and output handling both matter because sensitive data can leak through secondary systems; the test is necessary where a restricted assistant sometimes reveals private details in generated support logs.
Incorrect Answers
A:A scoped safety policy should prevent harmful responses while preserving legitimate help; this makes sense when a public service receives harassment and self-harm requests outside its permitted purpose, rather than the risk that a support log contains masked account information in normal use but exposes unmasked details on retry failures.
C:Layered controls prevent source text from overriding authority and block actual secret access; the trigger would be that a retrieved page tells an assistant to reveal a system secret during a normal summary, while the immediate question involves a support log contains masked account information in normal use but exposes unmasked details on retry failures.
D:No single natural-language instruction can replace controls at retrieval and action boundaries; the design helps when a security review finds the team relies only on a prompt asking the model to obey policy, but the present constraint is that a support log contains masked account information in normal use but exposes unmasked details on retry failures.