Question 610 A customer-support index contains an uploaded document with hidden HTML instructions asking the agent to email a password list. What prevention measure is most targeted to this underlying cause? Pin trusted tool schemas, audit every call and test malicious response payloads. Apply markup sanitation, prompt-injection scanning and least-privilege tools to ingestion and retrieval. Use source reputation, content review and quarantine workflows before indexing regulated guidance. Add moderation and evidence-quality checks to source onboarding and refreshed content. ExplanationCorrect Answer B:Hidden content can be indexed and reintroduced as apparent context; a tool call must remain separately authorized. Other Answers A:That hidden content can be indexed and reintroduced as apparent context; a tool call must remain separately authorized. C:Hidden content can be indexed and reintroduced as apparent context; a tool call must remain separately authorized. D:This control addresses a different exposure: A public forum archive mixes factual troubleshooting steps with abusive text, and the assistant reproduces insults from retrieved comments. It would not prevent this failure, because hidden content can be indexed and reintroduced as apparent context; a tool call must remain separately authorized.
Question 611 A public forum archive mixes factual troubleshooting steps with abusive text, and the assistant reproduces insults from retrieved comments. Which implementation decision best restores the intended behavior? Filter or quarantine untrusted scam pages using source-quality controls and restrict compliance retrieval to approved sources. Sanitize hidden markup, treat surviving document text as untrusted and block any unauthorized external email tool action. Verify contractual confidentiality and authorization before reusing the corpus; keep restricted contracts out of public retrieval scope. Curate authoritative technical segments and filter abusive passages while preserving relevant safe troubleshooting evidence. ExplanationCorrect Answer D:The action targets the stated cause: Context selection influences output style and safety; raw forum text may be unsuitable as model evidence. Check the repair by: run common troubleshooting questions and confirm the cited instructions remain useful without abusive language. Other Answers A:Context selection influences output style and safety; raw forum text may be unsuitable as model evidence. B:Context selection influences output style and safety; raw forum text may be unsuitable as model evidence. C:Context selection influences output style and safety; raw forum text may be unsuitable as model evidence.
Question 612 A public forum archive mixes factual troubleshooting steps with abusive text, and the assistant reproduces insults from retrieved comments. Which underlying technical principle explains why this occurs? Hidden content can be indexed and reintroduced as apparent context; a tool call must remain separately authorized. Corpus contamination can change generated recommendations even without direct prompt-injection commands. Permission to process a document internally does not automatically grant public redistribution rights or confidentiality waivers. Context selection influences output style and safety; raw forum text may be unsuitable as model evidence. ExplanationCorrect Answer D:Curate authoritative technical segments and filter abusive passages while preserving relevant safe troubleshooting evidence. Other Answers A:Context selection influences output style and safety; raw forum text may be unsuitable as model evidence. B:Context selection influences output style and safety; raw forum text may be unsuitable as model evidence. C:That mechanism describes another problem: An agent indexes internal vendor contracts for enterprise employees and then exposes the same retrieval index to public customers. It does not explain this case, where context selection influences output style and safety; raw forum text may be unsuitable as model evidence.
Question 613 A public forum archive mixes factual troubleshooting steps with abusive text, and the assistant reproduces insults from retrieved comments. After correction, what targeted test should demonstrate success? Run common troubleshooting questions and confirm the cited instructions remain useful without abusive language. Upload a synthetic malicious HTML file and verify no external email is sent while ordinary visible content remains searchable. Ask compliance benchmark questions and verify suspicious pages are excluded from retrieved evidence and final citations. Test a public account and confirm zero confidential contract passages appear in candidate search or output. ExplanationCorrect Answer A:Context selection influences output style and safety; raw forum text may be unsuitable as model evidence. Other Answers B:Run common troubleshooting questions and confirm the cited instructions remain useful without abusive language. C:Run common troubleshooting questions and confirm the cited instructions remain useful without abusive language. D:Run common troubleshooting questions and confirm the cited instructions remain useful without abusive language.
Question 614 A public forum archive mixes factual troubleshooting steps with abusive text, and the assistant reproduces insults from retrieved comments. Which release safeguard most directly addresses this failure class? Apply markup sanitation, prompt-injection scanning and least-privilege tools to ingestion and retrieval. Tag sources with allowed audiences and enforce audience-scoped retrieval paths at ingest and query time. Use source reputation, content review and quarantine workflows before indexing regulated guidance. Add moderation and evidence-quality checks to source onboarding and refreshed content. ExplanationCorrect Answer D:Context selection influences output style and safety; raw forum text may be unsuitable as model evidence. Other Answers A:Context selection influences output style and safety; raw forum text may be unsuitable as model evidence. B:This control addresses a different exposure: An agent indexes internal vendor contracts for enterprise employees and then exposes the same retrieval index to public customers. It would not prevent this failure, because context selection influences output style and safety; raw forum text may be unsuitable as model evidence. C:That context selection influences output style and safety; raw forum text may be unsuitable as model evidence.
Question 615 An ingestion process labels every external PDF trusted because the file is digitally signed, although it contains a malicious instruction to bypass business approvals. Which design change directly satisfies the stated constraints? Treat webpage text as untrusted evidence and enforce fixed trusted instructions and tool permissions regardless of embedded commands. Assess document content and originating authority separately from file integrity, then keep tool approvals enforced outside retrieval text. Sanitize hidden markup, treat surviving document text as untrusted and block any unauthorized external email tool action. Filter or quarantine untrusted scam pages using source-quality controls and restrict compliance retrieval to approved sources. ExplanationCorrect Answer B:The action targets the stated cause: A signature may establish file integrity or signer identity but does not elevate embedded prose into executable agent policy. Check the repair by: replay a signed malicious document and verify its instruction cannot bypass required action approval. Other Answers A:A signature may establish file integrity or signer identity but does not elevate embedded prose into executable agent policy. C:A signature may establish file integrity or signer identity but does not elevate embedded prose into executable agent policy. D:A signature may establish file integrity or signer identity but does not elevate embedded prose into executable agent policy.
Question 616 An ingestion process labels every external PDF trusted because the file is digitally signed, although it contains a malicious instruction to bypass business approvals. Which causal explanation best accounts for the failure? Prompt injection exploits confusion between retrieved content and application instructions; source text cannot escalate its own authority. Hidden content can be indexed and reintroduced as apparent context; a tool call must remain separately authorized. A signature may establish file integrity or signer identity but does not elevate embedded prose into executable agent policy. Corpus contamination can change generated recommendations even without direct prompt-injection commands. ExplanationCorrect Answer C:Assess document content and originating authority separately from file integrity, then keep tool approvals enforced outside retrieval text. Other Answers A:A signature may establish file integrity or signer identity but does not elevate embedded prose into executable agent policy. B:That mechanism describes another problem: A customer-support index contains an uploaded document with hidden HTML instructions asking the agent to email a password list. It does not explain this case, where a signature may establish file integrity or signer identity but does not elevate embedded prose into executable agent policy. D:A signature may establish file integrity or signer identity but does not elevate embedded prose into executable agent policy.
Question 617 Two language models achieve similar factual accuracy but one requires four times the compute cost per successful answer on the actual workload. Which engineering action most directly addresses this specific requirement? Monitor valid-request success, refusal rate and sampled answer quality alongside latency, segmented by prompt version. Measure factual support and numeric consistency alongside lexical overlap, emphasizing high-severity hallucination frequency. Compare quality-adjusted cost per successful response and task latency, not raw inference price alone. Evaluate the architecture on the target long-context task with citation correctness and context utilization metrics. ExplanationCorrect Answer C:The action targets the stated cause: Cost per token can mislead when models differ in retries, token length and correct-answer rate. Check the repair by: benchmark identical prompts with accuracy, tokens, retries and total cost for both models. Other Answers A:Cost per token can mislead when models differ in retries, token length and correct-answer rate. B:Cost per token can mislead when models differ in retries, token length and correct-answer rate. D:Cost per token can mislead when models differ in retries, token length and correct-answer rate.
Question 618 Two language models achieve similar factual accuracy but one requires four times the compute cost per successful answer on the actual workload. What architectural constraint is most directly responsible? Parameter count and general benchmark scores do not directly measure a system's ability to use distant evidence reliably. Speed can remain stable while application utility drops; response quality and reliability are separate production dimensions. Cost per token can mislead when models differ in retries, token length and correct-answer rate. N-gram similarity rewards shared wording and cannot prove generated facts are supported by source records. ExplanationCorrect Answer C:Compare quality-adjusted cost per successful response and task latency, not raw inference price alone. Other Answers A:Cost per token can mislead when models differ in retries, token length and correct-answer rate. B:Cost per token can mislead when models differ in retries, token length and correct-answer rate. D:That mechanism describes another problem: An evaluation report gives a high average BLEU score for summaries but hides hallucinated financial amounts in five important cases. It does not explain this case, where cost per token can mislead when models differ in retries, token length and correct-answer rate.
Question 619 Two language models achieve similar factual accuracy but one requires four times the compute cost per successful answer on the actual workload. What validation evidence would most specifically confirm the resolution? Benchmark identical prompts with accuracy, tokens, retries and total cost for both models. Compare generated numbers against cited source fields for every high-risk evaluation case. Chart correct resolution and refusal counts per deployed prompt revision and reproduce failing requests in offline tests. Place essential facts at varying document positions and record grounded answer performance for each position. ExplanationCorrect Answer A:Cost per token can mislead when models differ in retries, token length and correct-answer rate. Other Answers B:Benchmark identical prompts with accuracy, tokens, retries and total cost for both models. C:Benchmark identical prompts with accuracy, tokens, retries and total cost for both models. D:Benchmark identical prompts with accuracy, tokens, retries and total cost for both models.
Question 620 Two language models achieve similar factual accuracy but one requires four times the compute cost per successful answer on the actual workload. What prevention measure is most targeted to this underlying cause? Maintain separate severity-weighted factuality metrics and nonnegotiable high-risk acceptance thresholds. Track multi-objective efficiency metrics tied to a real task distribution when selecting production models. Define balanced quality and operational SLOs with version-attributed monitoring. Include task-specific context-position stress tests in model comparison. ExplanationCorrect Answer B:Cost per token can mislead when models differ in retries, token length and correct-answer rate. Other Answers A:Cost per token can mislead when models differ in retries, token length and correct-answer rate. C:This control addresses a different exposure: A deployed assistant responds quickly but increasingly refuses valid customer questions after a prompt rollout. It would not prevent this failure, because cost per token can mislead when models differ in retries, token length and correct-answer rate. D:That cost per token can mislead when models differ in retries, token length and correct-answer rate.
