Question 270
An agent can reach arbitrary external URLs through a general browsing connector. Which solution best resolves this operational gap?
- Constrain network destinations, tool arguments and egress privileges to approved endpoints.
- Mark retrieved text as untrusted data and isolate executable tool privileges.
- Enforce document and tool authorization at the retrieval boundary before generating answers.
- Check refund amount, caller scope and approvals in a backend policy gateway.
Explanation
Correct Answer
A:Network-level restrictions limit data exfiltration routes independent of the model's intent. Test approved network egress boundaries with adversarial cases and an application-enforced boundary; credentials and service policies still have to enforce approved network egress boundaries.
Incorrect Answers
B:Its main effect concerns untrusted-content isolation. The case explicitly hinges on approved network egress boundaries; to resolve it reliably, the implementation has to constrain network destinations, tool arguments and egress privileges to approved endpoints, which this choice does not establish.
C:The option might improve access enforcement ahead of inference; that remains insufficient for approved network egress boundaries. The case is resolved by approved network egress boundaries, which depends on a design that will constrain network destinations, tool arguments and egress privileges to approved endpoints.
D:The proposed action prioritizes server-side limits on financial actions. That can coexist with the right solution, but it cannot replace approved network egress boundaries: the application must still constrain network destinations, tool arguments and egress privileges to approved endpoints.
Question 271
A retrieval system filters classified documents only after a model has already summarized them. What design decision would best fit these facts?
- Enforce document and tool authorization at the retrieval boundary before generating answers.
- Minimize sensitive inputs and apply output checks for protected fields.
- Move sensitivity classification and entitlement checks before context injection.
- Mark retrieved text as untrusted data and isolate executable tool privileges.
Explanation
Correct Answer
C:Post-hoc output filtering cannot reverse the disclosure of restricted source data into inference. Test preventive filtering before model context with adversarial cases and an application-enforced boundary; credentials and service policies still have to enforce preventive filtering before model context.
Incorrect Answers
A:It is aimed at a different outcome, access enforcement ahead of inference. No evidence is given that it provides preventive filtering before model context; the needed mechanism is to move sensitivity classification and entitlement checks before context injection at the appropriate boundary.
B:It emphasizes multiple layers of sensitive-data protection without closing the stated gap in preventive filtering before model context. By contrast, the remedy to move sensitivity classification and entitlement checks before context injection acts on the condition that makes this scenario fail.
D:Improving untrusted-content isolation does not imply preventive filtering before model context. The selected design must move sensitivity classification and entitlement checks before context injection, whereas this choice omits that direct protection or verification step.
Question 272
A legal assistant produces a polished answer based on an expired statute and has no source-age check. Which intervention most directly addresses the failure?
- Include citation-faithfulness failures in the hazard analysis and test plan.
- Treat unsupported numerical interpolation as a high-severity hallucination requiring deterministic verification.
- Identify accountability and appeal risk before production release.
- Identify stale-evidence risk and add validity-date controls to retrieval.
Explanation
Correct Answer
D:Fluency can conceal obsolete grounding; temporal applicability must be part of the risk model. Assess threats to exposure to outdated authoritative sources by consequence, detectability and exploit path; an apparently successful answer can still hide failure to deliver exposure to outdated authoritative sources.
Incorrect Answers
A:The proposed measure is aimed at false grounding and citation laundering risk, not exposure to outdated authoritative sources. The architecture must instead identify stale-evidence risk and add validity-date controls to retrieval; a separate change is still needed for exposure to outdated authoritative sources.
B:The relevant trade-off is not primarily fabricated quantitative evidence risk. A successful fix needs exposure to outdated authoritative sources; otherwise exposure to outdated authoritative sources remains unverified even after this change.
C:The control would be justified if missing appeal and remediation process were the limiting factor. In this scenario the decisive target is exposure to outdated authoritative sources; success requires an architecture that will identify stale-evidence risk and add validity-date controls to retrieval.
Question 273
A financial agent sometimes invents decimal values absent from its tool result. What should the solution architect prioritize?
- Prioritize cross-tenant action risk using severity and privilege boundaries, not raw frequency.
- Include retention and secondary-use exposure in its risk assessment.
- Identify stale-evidence risk and add validity-date controls to retrieval.
- Treat unsupported numerical interpolation as a high-severity hallucination requiring deterministic verification.
- Model mid-execution authorization drift and require rechecks before sensitive commits.
Explanation
Correct Answer
D:A believable numeric detail can directly alter financial decisions even when the surrounding text is correct. Assess threats to fabricated quantitative evidence risk by consequence, detectability and exploit path; an apparently successful answer can still hide failure to deliver fabricated quantitative evidence risk.
Incorrect Answers
A:The proposed action prioritizes severity-sensitive tenant isolation risk. That can coexist with the right solution, but it cannot replace fabricated quantitative evidence risk: the application must still treat unsupported numerical interpolation as a high-severity hallucination requiring deterministic verification.
B:The design concern is real but different: operational data retention risk. It does not impose the specific requirement for fabricated quantitative evidence risk; a suitable solution must treat unsupported numerical interpolation as a high-severity hallucination requiring deterministic verification and demonstrate the result.
C:This addresses exposure to outdated authoritative sources; however, the problem asks for fabricated quantitative evidence risk. Without the safeguard to treat unsupported numerical interpolation as a high-severity hallucination requiring deterministic verification, the original failure mode is still possible.
E:There is a legitimate use case for revoked authority during autonomous execution, but it does not supply the test or enforcement needed for fabricated quantitative evidence risk. The design needs to treat unsupported numerical interpolation as a high-severity hallucination requiring deterministic verification.
Question 274
An automated approval workflow fails rarely but occasionally posts actions to the wrong customer tenant. Which architecture adjustment is most appropriate?
- Prioritize cross-tenant action risk using severity and privilege boundaries, not raw frequency.
- Model mid-execution authorization drift and require rechecks before sensitive commits.
- Include retention and secondary-use exposure in its risk assessment.
- Identify stale-evidence risk and add validity-date controls to retrieval.
Explanation
Correct Answer
A:Low-incidence access violations can outweigh common cosmetic answer mistakes in expected harm. Assess threats to severity-sensitive tenant isolation risk by consequence, detectability and exploit path; an apparently successful answer can still hide failure to deliver severity-sensitive tenant isolation risk.
Incorrect Answers
B:This recommendation targets revoked authority during autonomous execution at a different point in the workflow. It leaves severity-sensitive tenant isolation risk outside the design contract and does not require the team to prioritize cross-tenant action risk using severity and privilege boundaries, not raw frequency.
C:This is a valid architectural concern when optimizing operational data retention risk. Here it leaves severity-sensitive tenant isolation risk unaddressed, whereas the intended change must prioritize cross-tenant action risk using severity and privilege boundaries, not raw frequency before the case can be considered resolved.
D:This would make sense as an additional project activity for exposure to outdated authoritative sources. The priority required here is severity-sensitive tenant isolation risk, so the workflow still needs a control that will prioritize cross-tenant action risk using severity and privilege boundaries, not raw frequency.
Question 275
A model evaluates job applicants using historical labels that encode demographic disparities. Which next step is most defensible?
- Treat unsupported numerical interpolation as a high-severity hallucination requiring deterministic verification.
- Identify accountability and appeal risk before production release.
- Audit data, outcomes and subgroup disparities before relying on automated ranking.
- Model mid-execution authorization drift and require rechecks before sensitive commits.
Explanation
Correct Answer
C:Historical performance labels may reproduce inequitable treatment even when no protected field is explicitly requested.
Incorrect Answers
A:The option belongs to the broader work on fabricated quantitative evidence risk, but it cannot by itself deliver bias propagation from training labels. A decision on this case needs the team to audit data, outcomes and subgroup disparities before relying on automated ranking and verify that outcome.
B:This shifts attention toward missing appeal and remediation process when the relevant decision is about bias propagation from training labels. Unless the architecture can audit data, outcomes and subgroup disparities before relying on automated ranking, the failure described in the stem remains unresolved.
D:Although revoked authority during autonomous execution can be valuable, it is not the acceptance condition in the question. The missing provision is bias propagation from training labels, addressed by the control to audit data, outcomes and subgroup disparities before relying on automated ranking.
Question 276
A tool-using agent continues acting after a user's role is revoked mid-run. Which approach should be recommended?
- Model mid-execution authorization drift and require rechecks before sensitive commits.
- Treat unsupported numerical interpolation as a high-severity hallucination requiring deterministic verification.
- Prioritize cross-tenant action risk using severity and privilege boundaries, not raw frequency.
- Monitor both under-refusal harms and false refusals after guardrail changes.
Explanation
Correct Answer
A:An initially valid delegation can expire while a long-running workflow is still active. Assess threats to revoked authority during autonomous execution by consequence, detectability and exploit path; an apparently successful answer can still hide failure to deliver revoked authority during autonomous execution.
Incorrect Answers
B:There is a legitimate use case for fabricated quantitative evidence risk, but it does not supply the test or enforcement needed for revoked authority during autonomous execution. The design needs to model mid-execution authorization drift and require rechecks before sensitive commits.
C:Consider its actual operating scope: severity-sensitive tenant isolation risk. The requested assurance is revoked authority during autonomous execution, which calls for an implementation that can model mid-execution authorization drift and require rechecks before sensitive commits, not merely for a related improvement.
D:The alternative could reduce risk associated with safety regression hidden by usability gains, yet the described behavior requires revoked authority during autonomous execution. That outcome depends on the capability to model mid-execution authorization drift and require rechecks before sensitive commits.
Question 277
A prompt injection makes a support bot obey a forged system directive included in a knowledge article. What is the best production design response?
- Include citation-faithfulness failures in the hazard analysis and test plan.
- Monitor both under-refusal harms and false refusals after guardrail changes.
- Treat instruction laundering through retrieved content as a distinct adversarial threat.
- Evaluate operational and logging paths as part of the attack surface.
Explanation
Correct Answer
C:The unsafe behavior arises from source text being promoted to authority, not from ordinary factual error. Assess threats to prompt-injection trust-boundary exposure by consequence, detectability and exploit path; an apparently successful answer can still hide failure to deliver prompt-injection trust-boundary exposure.
Incorrect Answers
A:This would make sense as an additional project activity for false grounding and citation laundering risk. The priority required here is prompt-injection trust-boundary exposure, so the workflow still needs a control that will treat instruction laundering through retrieved content as a distinct adversarial threat.
B:By itself, this improves safety regression hidden by usability gains, but the case is not asking the architect to optimize that dimension. The requirement is prompt-injection trust-boundary exposure, supported by an approach to treat instruction laundering through retrieved content as a distinct adversarial threat.
D:The option might improve non-model information disclosure paths; that remains insufficient for prompt-injection trust-boundary exposure. The case is resolved by prompt-injection trust-boundary exposure, which depends on a design that will treat instruction laundering through retrieved content as a distinct adversarial threat.
Question 278
A chatbot refuses dangerous instructions correctly but exposes internal logs through an error response. Which change most directly satisfies this requirement?
- Treat instruction laundering through retrieved content as a distinct adversarial threat.
- Evaluate operational and logging paths as part of the attack surface.
- Monitor both under-refusal harms and false refusals after guardrail changes.
- Include retention and secondary-use exposure in its risk assessment.
Explanation
Correct Answer
B:Safe generated text does not mean application infrastructure protects sensitive diagnostics. Assess threats to non-model information disclosure paths by consequence, detectability and exploit path; an apparently successful answer can still hide failure to deliver non-model information disclosure paths.
Incorrect Answers
A:Although prompt-injection trust-boundary exposure can be valuable, it is not the acceptance condition in the question. The missing provision is non-model information disclosure paths, addressed by the control to evaluate operational and logging paths as part of the attack surface.
C:It emphasizes safety regression hidden by usability gains without closing the stated gap in non-model information disclosure paths. By contrast, the remedy to evaluate operational and logging paths as part of the attack surface acts on the condition that makes this scenario fail.
D:That is primarily a technique for operational data retention risk. It lacks the explicit measure for non-model information disclosure paths that is necessary here, namely a system that will evaluate operational and logging paths as part of the attack surface.
Question 279
A summarization service's end-to-end success metric excludes all requests that timed out. Which solution best resolves this operational gap?
- Identify accountability and appeal risk before production release.
- Count failed and abandoned work in reliability and risk reporting.
- Include retention and secondary-use exposure in its risk assessment.
- Monitor both under-refusal harms and false refusals after guardrail changes.
Explanation
Correct Answer
B:Selection bias toward completed jobs hides operational failure from governance dashboards. Assess threats to honest measurement of silent timeouts by consequence, detectability and exploit path; an apparently successful answer can still hide failure to deliver honest measurement of silent timeouts.
Incorrect Answers
A:The alternative could reduce risk associated with missing appeal and remediation process, yet the described behavior requires honest measurement of silent timeouts. That outcome depends on the capability to count failed and abandoned work in reliability and risk reporting.
C:The relevant trade-off is not primarily operational data retention risk. A successful fix needs honest measurement of silent timeouts; otherwise honest measurement of silent timeouts remains unverified even after this change.
D:The control would be justified if safety regression hidden by usability gains were the limiting factor. In this scenario the decisive target is honest measurement of silent timeouts; success requires an architecture that will count failed and abandoned work in reliability and risk reporting.
Question 280
A regulatory team accepts a model's cited answer without checking whether cited pages support the claims. What design decision would best fit these facts?
- Identify stale-evidence risk and add validity-date controls to retrieval.
- Treat instruction laundering through retrieved content as a distinct adversarial threat.
- Model mid-execution authorization drift and require rechecks before sensitive commits.
- Include citation-faithfulness failures in the hazard analysis and test plan.
- Monitor both under-refusal harms and false refusals after guardrail changes.
Explanation
Correct Answer
D:A reference marker may look reassuring while pointing to irrelevant evidence. Assess threats to false grounding and citation laundering risk by consequence, detectability and exploit path; an apparently successful answer can still hide failure to deliver false grounding and citation laundering risk.
Incorrect Answers
A:The option might improve exposure to outdated authoritative sources; that remains insufficient for false grounding and citation laundering risk. The case is resolved by false grounding and citation laundering risk, which depends on a design that will include citation-faithfulness failures in the hazard analysis and test plan.
B:The proposed action prioritizes prompt-injection trust-boundary exposure. That can coexist with the right solution, but it cannot replace false grounding and citation laundering risk: the application must still include citation-faithfulness failures in the hazard analysis and test plan.
C:The design concern is real but different: revoked authority during autonomous execution. It does not impose the specific requirement for false grounding and citation laundering risk; a suitable solution must include citation-faithfulness failures in the hazard analysis and test plan and demonstrate the result.
E:That choice would be responsive to safety regression hidden by usability gains, but the reported defect is different. It leaves false grounding and citation laundering risk unproved because there is no commitment to include citation-faithfulness failures in the hazard analysis and test plan.