Question 290 A support bot exposes the full credit-card number in MLflow traces while producing a redacted response to the customer. Which verification is strongest for this particular failure mode? Compare response correctness on a labeled Indonesian set and document coverage gaps from the model card. Inject the malicious paragraph in a test corpus and confirm no network tool receives credential values. Inject synthetic card data and verify neither trace attributes nor model outputs contain the raw identifier. Run ambiguous clinical prompts and verify the system declines unsupported prescribing while providing appropriate general guidance. ExplanationCorrect Answer C:Output masking does not protect logs or tool traces; leakage can occur earlier in the data path. Other Answers A:Inject synthetic card data and verify neither trace attributes nor model outputs contain the raw identifier. B:Inject synthetic card data and verify neither trace attributes nor model outputs contain the raw identifier. D:Inject synthetic card data and verify neither trace attributes nor model outputs contain the raw identifier.
Question 291 A support bot exposes the full credit-card number in MLflow traces while producing a redacted response to the customer. What change to engineering governance would reduce this risk over time? Maintain high-risk domain policies and independently review clinical safety evaluation cases. Apply end-to-end data minimization and log-scrubbing checks to all agent components. Enforce least-privilege tools and continuous prompt-injection adversarial tests. Maintain target-language evidence requirements and uncertainty flags during model selection. ExplanationCorrect Answer B:Output masking does not protect logs or tool traces; leakage can occur earlier in the data path. Other Answers A:This control addresses a different exposure: A medical Q&A assistant is asked for a diagnosis despite incomplete records, and it confidently suggests a specific prescription. It would not prevent this failure, because output masking does not protect logs or tool traces; leakage can occur earlier in the data path. C:That output masking does not protect logs or tool traces; leakage can occur earlier in the data path. D:Output masking does not protect logs or tool traces; leakage can occur earlier in the data path.
Question 292 An assistant's final answer is safe, but its tool-call arguments contain an unauthorized delete operation generated from a user's “clean up” prompt. Which engineering action most directly addresses this specific requirement? Apply scope guardrails requiring educational information, uncertainty and referral for clinical decisions, and avoid actionable patient-specific prescriptions. Gate side-effect tools with explicit allowed actions, resource-level authorization and confirmation before the call executes. Require use of supplied specification fields, forbid unsupported values and include an explicit not_found response for missing attributes. Ignore document-embedded commands, constrain tool calls to server-side permissions and redact credentials from the model context. ExplanationCorrect Answer B:The action targets the stated cause: Final-response safety filters run too late to prevent a destructive tool invocation that has already occurred. Check the repair by: replay a destructive prompt and inspect that the attempted tool call is blocked before execution and recorded for review. Other Answers A:Final-response safety filters run too late to prevent a destructive tool invocation that has already occurred. C:Final-response safety filters run too late to prevent a destructive tool invocation that has already occurred. D:Final-response safety filters run too late to prevent a destructive tool invocation that has already occurred.
Question 293 An assistant's final answer is safe, but its tool-call arguments contain an unauthorized delete operation generated from a user's “clean up” prompt. What architectural constraint is most directly responsible? Final-response safety filters run too late to prevent a destructive tool invocation that has already occurred. Retrieved passages are untrusted task data, not higher-priority instructions authorizing external actions. Generative confidence is not clinical verification; missing patient context makes definitive treatment unsafe. General prompts allow parametric guesses to compete with retrieved evidence; instruction hierarchy must make source-grounded facts decisive. ExplanationCorrect Answer A:Gate side-effect tools with explicit allowed actions, resource-level authorization and confirmation before the call executes. Other Answers B:That mechanism describes another problem: A retrieved contract includes a paragraph instructing an agent to send the customer's credentials to a web endpoint. It does not explain this case, where final-response safety filters run too late to prevent a destructive tool invocation that has already occurred. C:Final-response safety filters run too late to prevent a destructive tool invocation that has already occurred. D:Final-response safety filters run too late to prevent a destructive tool invocation that has already occurred.
Question 294 An assistant's final answer is safe, but its tool-call arguments contain an unauthorized delete operation generated from a user's “clean up” prompt. What validation evidence would most specifically confirm the resolution? Inject the malicious paragraph in a test corpus and confirm no network tool receives credential values. Test deliberately absent specifications and verify the answer says unavailable instead of supplying invented measurements. Run ambiguous clinical prompts and verify the system declines unsupported prescribing while providing appropriate general guidance. Replay a destructive prompt and inspect that the attempted tool call is blocked before execution and recorded for review. ExplanationCorrect Answer D:Final-response safety filters run too late to prevent a destructive tool invocation that has already occurred. Other Answers A:Replay a destructive prompt and inspect that the attempted tool call is blocked before execution and recorded for review. B:Replay a destructive prompt and inspect that the attempted tool call is blocked before execution and recorded for review. C:Replay a destructive prompt and inspect that the attempted tool call is blocked before execution and recorded for review.
Question 295 An assistant's final answer is safe, but its tool-call arguments contain an unauthorized delete operation generated from a user's “clean up” prompt. What prevention measure is most targeted to this underlying cause? Separate read and write permissions and require server-side confirmation policies for destructive operations. Enforce least-privilege tools and continuous prompt-injection adversarial tests. Measure groundedness on missing-field and conflicting-source examples during every prompt revision. Maintain high-risk domain policies and independently review clinical safety evaluation cases. ExplanationCorrect Answer A:Final-response safety filters run too late to prevent a destructive tool invocation that has already occurred. Other Answers B:That final-response safety filters run too late to prevent a destructive tool invocation that has already occurred. C:Final-response safety filters run too late to prevent a destructive tool invocation that has already occurred. D:This control addresses a different exposure: A medical Q&A assistant is asked for a diagnosis despite incomplete records, and it confidently suggests a specific prescription. It would not prevent this failure, because final-response safety filters run too late to prevent a destructive tool invocation that has already occurred.
Question 296 A team selects the largest available model for short classification messages, producing accurate labels but unacceptable latency and expense. Which implementation decision best restores the intended behavior? Reject the incompatible model and choose an adequately performing candidate with a license permitting the intended commercial use. Choose a model supporting the required context and evaluate citation accuracy under measured full-context loads or redesign retrieval to fit. Select or route to a candidate validated on the actual target languages and legal terminology, even if its unit inference cost is higher. Benchmark smaller candidate models for classification accuracy and P95 latency, selecting the least costly model meeting both targets. ExplanationCorrect Answer D:The action targets the stated cause: Model size does not determine whether a simple constrained task needs expensive general reasoning capacity. Check the repair by: run labeled classification and load tests with cost per thousand requests, macro-F1 and P95 latency. Other Answers A:Model size does not determine whether a simple constrained task needs expensive general reasoning capacity. B:Model size does not determine whether a simple constrained task needs expensive general reasoning capacity. C:Model size does not determine whether a simple constrained task needs expensive general reasoning capacity.
Question 297 A team selects the largest available model for short classification messages, producing accurate labels but unacceptable latency and expense. Which underlying technical principle explains why this occurs? Model size does not determine whether a simple constrained task needs expensive general reasoning capacity. A model's advertised context window constrains input and output budgeting; silently truncated evidence breaks grounding. A model card's license and permitted uses constrain deployment independently of benchmark quality. Model language coverage and domain performance vary; English-only scores do not predict accuracy on Japanese legal text. ExplanationCorrect Answer A:Benchmark smaller candidate models for classification accuracy and P95 latency, selecting the least costly model meeting both targets. Other Answers B:Model size does not determine whether a simple constrained task needs expensive general reasoning capacity. C:Model size does not determine whether a simple constrained task needs expensive general reasoning capacity. D:That mechanism describes another problem: A multilingual legal assistant's inexpensive model handles English but fails on Japanese clauses that affect customer obligations. It does not explain this case, where model size does not determine whether a simple constrained task needs expensive general reasoning capacity.
Question 298 A team selects the largest available model for short classification messages, producing accurate labels but unacceptable latency and expense. After correction, what targeted test should demonstrate success? Count real prompt tokens on worst-case policies and verify final citations reference included—not truncated—text. Check the exact published license terms and record the selection decision alongside measured task performance. Compare clause-extraction errors by language and ensure the selected model meets each locale's safety threshold. Run labeled classification and load tests with cost per thousand requests, macro-F1 and P95 latency. ExplanationCorrect Answer D:Model size does not determine whether a simple constrained task needs expensive general reasoning capacity. Other Answers A:Run labeled classification and load tests with cost per thousand requests, macro-F1 and P95 latency. B:Run labeled classification and load tests with cost per thousand requests, macro-F1 and P95 latency. C:Run labeled classification and load tests with cost per thousand requests, macro-F1 and P95 latency.
Question 299 A team selects the largest available model for short classification messages, producing accurate labels but unacceptable latency and expense. Which release safeguard most directly addresses this failure class? Require license and permitted-use review in every model-catalog promotion checklist. Maintain language-segmented evaluation and safe fallback for unsupported locales. Set maximum-context validation and retrieval fallback rules in the model-selection contract. Use task-specific quality, cost and latency gates instead of model-size preference. ExplanationCorrect Answer D:Model size does not determine whether a simple constrained task needs expensive general reasoning capacity. Other Answers A:Model size does not determine whether a simple constrained task needs expensive general reasoning capacity. B:This control addresses a different exposure: A multilingual legal assistant's inexpensive model handles English but fails on Japanese clauses that affect customer obligations. It would not prevent this failure, because model size does not determine whether a simple constrained task needs expensive general reasoning capacity. C:That model size does not determine whether a simple constrained task needs expensive general reasoning capacity.
Question 300 A multilingual legal assistant's inexpensive model handles English but fails on Japanese clauses that affect customer obligations. Which design change directly satisfies the stated constraints? Treat the benchmark as insufficient, run an Indonesian task evaluation and prefer a model with proven target-language behavior. Select or route to a candidate validated on the actual target languages and legal terminology, even if its unit inference cost is higher. Choose a model supporting the required context and evaluate citation accuracy under measured full-context loads or redesign retrieval to fit. Benchmark smaller candidate models for classification accuracy and P95 latency, selecting the least costly model meeting both targets. ExplanationCorrect Answer B:The action targets the stated cause: Model language coverage and domain performance vary; English-only scores do not predict accuracy on Japanese legal text. Check the repair by: compare clause-extraction errors by language and ensure the selected model meets each locale's safety threshold. Other Answers A:Model language coverage and domain performance vary; English-only scores do not predict accuracy on Japanese legal text. C:Model language coverage and domain performance vary; English-only scores do not predict accuracy on Japanese legal text. D:Model language coverage and domain performance vary; English-only scores do not predict accuracy on Japanese legal text.
