100% Latest & Updated Microsoft AB-100 Practice Test Questions, Exam Dumps & Verified Answers!
30 Days Free Updates, Instant Download!
Question 370
A model-tuning experiment includes raw employee salaries without an approved data use purpose. Which specific remediation should the business adopt?
- Deidentify test and evaluation datasets and scope access to reviewers.
- Exclude or minimize protected data until lawful purpose and access authorization are established.
- Validate data lineage and inspect anomalous training inputs before model promotion.
- Preserve source classification and apply output handling controls to retrieved fragments.
Explanation
Correct Answer
B:Fine-tuning use needs its own permitted basis; access for operations is not automatic consent for training.
Incorrect Answers
A:Evaluation does not require unrestricted production PII exposure, whereas fine-tuning use needs its own permitted basis; access for operations is not automatic consent for training.
C:Poisoning defense requires trust in the data pipeline, not just final accuracy tests, but fine-tuning use needs its own permitted basis; access for operations is not automatic consent for training.
D:Security labels must follow sensitive content through the generated response path. Fine-tuning use needs its own permitted basis; access for operations is not automatic consent for training.
Question 371
A service bot stores customer identifiers in evaluation examples for unrelated developers. Which solution approach best meets the stated requirement?
- Apply location and minimization controls to telemetry and audit storage.
- Enforce document-level access control during retrieval for every caller.
- Deidentify test and evaluation datasets and scope access to reviewers.
- Preserve source classification and apply output handling controls to retrieved fragments.
Explanation
Correct Answer
C:Evaluation does not require unrestricted production PII exposure.
Incorrect Answers
A:Observability data may contain protected information and require the same governance review. Evaluation does not require unrestricted production PII exposure.
B:Indexing does not erase the underlying source permissions, but evaluation does not require unrestricted production PII exposure.
D:Security labels must follow sensitive content through the generated response path; evaluation does not require unrestricted production PII exposure.
Question 372
A retrieval system returns a document fragment but its classification label is lost before the answer is shown. Which option most directly closes the documented gap?
- Enforce document-level access control during retrieval for every caller.
- Exclude or minimize protected data until lawful purpose and access authorization are established.
- Maintain an agent inventory with data sources, tools, privileges and classification.
- Preserve source classification and apply output handling controls to retrieved fragments.
Explanation
Correct Answer
D:Security labels must follow sensitive content through the generated response path.
Incorrect Answers
A:Indexing does not erase the underlying source permissions, but security labels must follow sensitive content through the generated response path.
B:Fine-tuning use needs its own permitted basis; access for operations is not automatic consent for training. Security labels must follow sensitive content through the generated response path.
C:An inventory enables risk review of the organization’s actual AI attack surface; security labels must follow sensitive content through the generated response path.
Question 373
A regulator asks which employee approved an autonomous refund and which model recommendation was shown. Which concrete solution decision should the team make?
- Produce protected residency, configuration and execution audit records.
- Log the approver, recommendation, evidence and final executed transaction ID.
- Capture authorized retrieval events with subject, source and purpose metadata.
- Record prompt, model and knowledge versions for the execution trace.
Explanation
Correct Answer
B:A linked record distinguishes AI advice from the actual accountable human action.
Incorrect Answers
A:Compliance evidence must tie the specific transaction to effective controls and versions. A linked record distinguishes AI advice from the actual accountable human action.
C:Access auditing can demonstrate when protected information was used by the assistant, whereas a linked record distinguishes AI advice from the actual accountable human action.
D:Run provenance supports investigation of behavior at a specific time. A linked record distinguishes AI advice from the actual accountable human action.
Question 374
A model output changed after a prompt update but no one knows which prompt produced a disputed answer. Which design decision should the architect prioritize?
- Propagate a stable cross-system correlation ID through agent and ERP actions.
- Review model provenance, license and integrity before approving deployment.
- Record prompt, model and knowledge versions for the execution trace.
- Capture authorized retrieval events with subject, source and purpose metadata.
Explanation
Correct Answer
C:Run provenance supports investigation of behavior at a specific time.
Incorrect Answers
A:Transaction linkage supports accountability across the agent and business record; run provenance supports investigation of behavior at a specific time.
B:Third-party model supply-chain checks reduce security and legal uncertainty, whereas run provenance supports investigation of behavior at a specific time.
D:Access auditing can demonstrate when protected information was used by the assistant, but run provenance supports investigation of behavior at a specific time.
Question 375
An agent reads a restricted customer file and produces a summary; administrators need a trace of access. Which controlled change is most appropriate for this situation?
- Propagate a stable cross-system correlation ID through agent and ERP actions.
- Log the approver, recommendation, evidence and final executed transaction ID.
- Produce protected residency, configuration and execution audit records.
- Capture authorized retrieval events with subject, source and purpose metadata.
Explanation
Correct Answer
D:Access auditing can demonstrate when protected information was used by the assistant.
Incorrect Answers
A:Transaction linkage supports accountability across the agent and business record, whereas access auditing can demonstrate when protected information was used by the assistant.
B:A linked record distinguishes AI advice from the actual accountable human action. Access auditing can demonstrate when protected information was used by the assistant.
C:Compliance evidence must tie the specific transaction to effective controls and versions, but access auditing can demonstrate when protected information was used by the assistant.
Question 376
Incident investigators cannot correlate a generated purchase request with the actual ERP transaction record. Which design choice targets the observed problem?
- Propagate a stable cross-system correlation ID through agent and ERP actions.
- Revoke the token immediately, contain access and audit recent transactions.
- Log the approver, recommendation, evidence and final executed transaction ID.
- Record prompt, model and knowledge versions for the execution trace.
Explanation
Correct Answer
A:Transaction linkage supports accountability across the agent and business record.
Incorrect Answers
B:Credential incident containment requires prompt revocation and impact reconstruction, but transaction linkage supports accountability across the agent and business record.
C:A linked record distinguishes AI advice from the actual accountable human action. Transaction linkage supports accountability across the agent and business record.
D:Run provenance supports investigation of behavior at a specific time; transaction linkage supports accountability across the agent and business record.
Question 377
A team wants to deploy an externally supplied model whose provenance and license terms are unknown. Which architecture response most directly addresses the risk?
- Assess and constrain data flows across all integrated external services.
- Validate data lineage and inspect anomalous training inputs before model promotion.
- Review model provenance, license and integrity before approving deployment.
- Revoke exposed credentials and use secure secret management with minimum scoped access.
Explanation
Correct Answer
C:Third-party model supply-chain checks reduce security and legal uncertainty.
Incorrect Answers
A:Cross-service tool calls can cause regulated transfers even when core hosting is compliant, but third-party model supply-chain checks reduce security and legal uncertainty.
B:Poisoning defense requires trust in the data pipeline, not just final accuracy tests. Third-party model supply-chain checks reduce security and legal uncertainty.
D:Credential exposure can enable unauthorized inference and data access, whereas third-party model supply-chain checks reduce security and legal uncertainty.
Question 378
A shared model endpoint is exposed with an unrestricted key committed to source control. Which recommendation addresses the identified failure most directly?
- Validate data lineage and inspect anomalous training inputs before model promotion.
- Revoke the token immediately, contain access and audit recent transactions.
- Inventory affected deployments and perform controlled replacement with regression checks.
- Revoke exposed credentials and use secure secret management with minimum scoped access.
Explanation
Correct Answer
D:Credential exposure can enable unauthorized inference and data access.
Incorrect Answers
A:Poisoning defense requires trust in the data pipeline, not just final accuracy tests. Credential exposure can enable unauthorized inference and data access.
B:Credential incident containment requires prompt revocation and impact reconstruction; credential exposure can enable unauthorized inference and data access.
C:Known model vulnerabilities require traceable remediation of the actual deployed versions, whereas credential exposure can enable unauthorized inference and data access.
Question 379
A custom model has a suspicious data poisoning pattern in new tuning examples. What should the solution architect recommend for this specific need?
- Validate data lineage and inspect anomalous training inputs before model promotion.
- Review model provenance, license and integrity before approving deployment.
- Inventory affected deployments and perform controlled replacement with regression checks.
- Evaluate retrieval poisoning risk and apply source trust/approval controls.
Explanation
Correct Answer
A:Poisoning defense requires trust in the data pipeline, not just final accuracy tests.
Incorrect Answers
B:Third-party model supply-chain checks reduce security and legal uncertainty, whereas poisoning defense requires trust in the data pipeline, not just final accuracy tests.
C:Known model vulnerabilities require traceable remediation of the actual deployed versions. Poisoning defense requires trust in the data pipeline, not just final accuracy tests.
D:External content can be altered to influence agent answers and tool choices, but poisoning defense requires trust in the data pipeline, not just final accuracy tests.
Question 380
A model provider announces a security issue affecting one deployed model version. What is the most defensible intervention for this case?
- Review model provenance, license and integrity before approving deployment.
- Inventory affected deployments and perform controlled replacement with regression checks.
- Preserve source classification and apply output handling controls to retrieved fragments.
- Revoke exposed credentials and use secure secret management with minimum scoped access.
Explanation
Correct Answer
B:Known model vulnerabilities require traceable remediation of the actual deployed versions.
Incorrect Answers
A:Third-party model supply-chain checks reduce security and legal uncertainty, whereas known model vulnerabilities require traceable remediation of the actual deployed versions.
C:Security labels must follow sensitive content through the generated response path, but known model vulnerabilities require traceable remediation of the actual deployed versions.
D:Credential exposure can enable unauthorized inference and data access. Known model vulnerabilities require traceable remediation of the actual deployed versions.
Agentic AI Business Solutions Architect (AB-100) Practice Tests