Microsoft AI-103 Responsible AI Controls And Agent Governance Practice Test

 

This practice topic covers Responsible AI Controls and Agent Governance for Microsoft AI-103. Questions are original and aligned to the current Microsoft skills outline. For broader exam preparation, review the Microsoft AI-103 Exam Dumps page.

Question 1

A workload has a specific constraint: for different risk entry points. The implementation must choose correctly between input and output moderation. Which option best satisfies the constraint?

1. Apply input moderation to unsafe requests before model execution and output moderation to generated-response risks; place the control at the actual risk entry point.
2. Rely on prompt instructions alone and omit an independent enforcement control.
3. Record the action only after execution and omit a pre-action approval for irreversible changes.
4. Trust a fluent response as proof that the output is grounded and safe.
5. Let retrieved or image-embedded instructions override the system policy.

Correct Answer: A

Correct Answer

Answer A is correct because this choice directly implements the required decision for input versus output moderation for different risk entry points. It addresses the scenario at the correct stage of the Azure AI solution.

Incorrect Answers

Answer B is incorrect because this choice would rely on prompt instructions alone and omit an independent enforcement control. That can address a neighboring concern, but it does not implement input versus output moderation for different risk entry points; the decisive requirement in the stem therefore remains unresolved.

Answer C is incorrect because this choice would record the action only after execution and omit a pre-action approval for irreversible changes. That can address a neighboring concern, but it does not implement input versus output moderation for different risk entry points; the decisive requirement in the stem therefore remains unresolved.

Answer D is incorrect because this choice would trust a fluent response as proof that the output is grounded and safe. That can address a neighboring concern, but it does not implement input versus output moderation for different risk entry points; the decisive requirement in the stem therefore remains unresolved.

Answer E is incorrect because this choice would let retrieved or image-embedded instructions override the system policy. That can address a neighboring concern, but it does not implement input versus output moderation for different risk entry points; the decisive requirement in the stem therefore remains unresolved.

 

Question 2

In a production AI solution, the workload operates under a stated false-positive tolerance. Which action should the engineer take to handle category threshold choice correctly?

1. Let retrieved or image-embedded instructions override the system policy.
2. Trust a fluent response as proof that the output is grounded and safe.
3. Record the action only after execution and omit a pre-action approval for irreversible changes.
4. Give the agent broad standing permissions so authorization failures cannot occur.
5. Use category threshold choice under a stated false-positive tolerance.

Correct Answer: E

Correct Answer

Answer E is correct because this choice directly implements the required decision for category threshold choice under a stated false-positive tolerance. It addresses the scenario at the correct stage of the Azure AI solution.

Incorrect Answers

Answer A is incorrect because this choice would let retrieved or image-embedded instructions override the system policy. That can address a neighboring concern, but it does not implement category threshold choice under a stated false-positive tolerance; the decisive requirement in the stem therefore remains unresolved.

Answer B is incorrect because this choice would trust a fluent response as proof that the output is grounded and safe. That can address a neighboring concern, but it does not implement category threshold choice under a stated false-positive tolerance; the decisive requirement in the stem therefore remains unresolved.

Answer C is incorrect because this choice would record the action only after execution and omit a pre-action approval for irreversible changes. That can address a neighboring concern, but it does not implement category threshold choice under a stated false-positive tolerance; the decisive requirement in the stem therefore remains unresolved.

Answer D is incorrect because this choice would give the agent broad standing permissions so authorization failures cannot occur. That can address a neighboring concern, but it does not implement category threshold choice under a stated false-positive tolerance; the decisive requirement in the stem therefore remains unresolved.

 

Question 3

A workload has a specific constraint: where policy specifies review. The implementation must choose correctly between blocking and annotating. Which option best satisfies the constraint?

1. Trust a fluent response as proof that the output is grounded and safe.
2. Let retrieved or image-embedded instructions override the system policy.
3. Rely on prompt instructions alone and omit an independent enforcement control.
4. Flag and route the result for review when the policy requires review rather than automatic blocking.
5. Record the action only after execution and omit a pre-action approval for irreversible changes.

Correct Answer: D

Correct Answer

Answer D is correct because this choice directly implements the required decision for blocking versus annotating where policy specifies review. It addresses the scenario at the correct stage of the Azure AI solution.

Incorrect Answers

Answer A is incorrect because this choice would trust a fluent response as proof that the output is grounded and safe. That can address a neighboring concern, but it does not implement blocking versus annotating where policy specifies review; the decisive requirement in the stem therefore remains unresolved.

Answer B is incorrect because this choice would let retrieved or image-embedded instructions override the system policy. That can address a neighboring concern, but it does not implement blocking versus annotating where policy specifies review; the decisive requirement in the stem therefore remains unresolved.

Answer C is incorrect because this choice would rely on prompt instructions alone and omit an independent enforcement control. That can address a neighboring concern, but it does not implement blocking versus annotating where policy specifies review; the decisive requirement in the stem therefore remains unresolved.

Answer E is incorrect because this choice would record the action only after execution and omit a pre-action approval for irreversible changes. That can address a neighboring concern, but it does not implement blocking versus annotating where policy specifies review; the decisive requirement in the stem therefore remains unresolved.

 

Question 4

A runtime design must distinguish prompt attack detection from harmful-content classification. Which implementation uses the correct capability for the requirement?

1. Give the agent broad standing permissions so authorization failures cannot occur.
2. Rely on prompt instructions alone and omit an independent enforcement control.
3. Use prompt-attack detection for instruction-manipulation attacks and harmful-content classification for unsafe content categories.
4. Suppress safety telemetry so the aggregate pass rate improves.
5. Trust a fluent response as proof that the output is grounded and safe.

Correct Answer: C

Correct Answer

Answer C is correct because this choice directly implements the required decision for prompt attack detection versus harmful-content classification. It addresses the scenario at the correct stage of the Azure AI solution.

Incorrect Answers

Answer A is incorrect because this choice would give the agent broad standing permissions so authorization failures cannot occur. That can address a neighboring concern, but it does not implement prompt attack detection versus harmful-content classification; the decisive requirement in the stem therefore remains unresolved.

Answer B is incorrect because this choice would rely on prompt instructions alone and omit an independent enforcement control. That can address a neighboring concern, but it does not implement prompt attack detection versus harmful-content classification; the decisive requirement in the stem therefore remains unresolved.

Answer D is incorrect because this choice would suppress safety telemetry so the aggregate pass rate improves. That can address a neighboring concern, but it does not implement prompt attack detection versus harmful-content classification; the decisive requirement in the stem therefore remains unresolved.

Answer E is incorrect because this choice would trust a fluent response as proof that the output is grounded and safe. That can address a neighboring concern, but it does not implement prompt attack detection versus harmful-content classification; the decisive requirement in the stem therefore remains unresolved.

 

Question 5

In a production AI solution, the control must act before a state-changing tool action. Which action should the engineer take to handle guardrail placement correctly?

1. Rely on prompt instructions alone and omit an independent enforcement control.
2. Use guardrail placement before a state-changing tool action.
3. Record the action only after execution and omit a pre-action approval for irreversible changes.
4. Give the agent broad standing permissions so authorization failures cannot occur.
5. Let retrieved or image-embedded instructions override the system policy.

Correct Answer: B

Correct Answer

Answer B is correct because this choice directly implements the required decision for guardrail placement before a state-changing tool action. It addresses the scenario at the correct stage of the Azure AI solution.

Incorrect Answers

Answer A is incorrect because this choice would rely on prompt instructions alone and omit an independent enforcement control. That can address a neighboring concern, but it does not implement guardrail placement before a state-changing tool action; the decisive requirement in the stem therefore remains unresolved.

Answer C is incorrect because this choice would record the action only after execution and omit a pre-action approval for irreversible changes. That can address a neighboring concern, but it does not implement guardrail placement before a state-changing tool action; the decisive requirement in the stem therefore remains unresolved.

Answer D is incorrect because this choice would give the agent broad standing permissions so authorization failures cannot occur. That can address a neighboring concern, but it does not implement guardrail placement before a state-changing tool action; the decisive requirement in the stem therefore remains unresolved.

Answer E is incorrect because this choice would let retrieved or image-embedded instructions override the system policy. That can address a neighboring concern, but it does not implement guardrail placement before a state-changing tool action; the decisive requirement in the stem therefore remains unresolved.

 

Question 6

In a production AI solution, the following condition occurs: an acceptable prompt produces unsafe output. Which action should the engineer take to handle risk-specific control correctly?

1. Apply a control that targets the observed safety risk rather than a generic quality check when an acceptable prompt produces unsafe output.
2. Let retrieved or image-embedded instructions override the system policy.
3. Rely on prompt instructions alone and omit an independent enforcement control.
4. Trust a fluent response as proof that the output is grounded and safe.
5. Give the agent broad standing permissions so authorization failures cannot occur.

Correct Answer: A

Correct Answer

Answer A is correct because this choice directly implements the required decision for risk-specific control when an acceptable prompt produces unsafe output. It addresses the scenario at the correct stage of the Azure AI solution.

Incorrect Answers

Answer B is incorrect because this choice would let retrieved or image-embedded instructions override the system policy. That can address a neighboring concern, but it does not implement risk-specific control when an acceptable prompt produces unsafe output; the decisive requirement in the stem therefore remains unresolved.

Answer C is incorrect because this choice would rely on prompt instructions alone and omit an independent enforcement control. That can address a neighboring concern, but it does not implement risk-specific control when an acceptable prompt produces unsafe output; the decisive requirement in the stem therefore remains unresolved.

Answer D is incorrect because this choice would trust a fluent response as proof that the output is grounded and safe. That can address a neighboring concern, but it does not implement risk-specific control when an acceptable prompt produces unsafe output; the decisive requirement in the stem therefore remains unresolved.

Answer E is incorrect because this choice would give the agent broad standing permissions so authorization failures cannot occur. That can address a neighboring concern, but it does not implement risk-specific control when an acceptable prompt produces unsafe output; the decisive requirement in the stem therefore remains unresolved.

 

Question 7

In a production AI solution, the following condition occurs: the safety service is unavailable. Which action should the engineer take to handle failure behavior correctly?

1. Suppress safety telemetry so the aggregate pass rate improves.
2. Trust a fluent response as proof that the output is grounded and safe.
3. Rely on prompt instructions alone and omit an independent enforcement control.
4. Block every request in the category even though the policy requires differentiated handling.
5. Fail safely or route the request for review when the mandatory safety service is unavailable.

Correct Answer: E

Correct Answer

Answer E is correct because this choice directly implements the required decision for failure behavior when the safety service is unavailable. It addresses the scenario at the correct stage of the Azure AI solution.

Incorrect Answers

Answer A is incorrect because this choice would suppress safety telemetry so the aggregate pass rate improves. That can address a neighboring concern, but it does not implement failure behavior when the safety service is unavailable; the decisive requirement in the stem therefore remains unresolved.

Answer B is incorrect because this choice would trust a fluent response as proof that the output is grounded and safe. That can address a neighboring concern, but it does not implement failure behavior when the safety service is unavailable; the decisive requirement in the stem therefore remains unresolved.

Answer C is incorrect because this choice would rely on prompt instructions alone and omit an independent enforcement control. That can address a neighboring concern, but it does not implement failure behavior when the safety service is unavailable; the decisive requirement in the stem therefore remains unresolved.

Answer D is incorrect because this choice would block every request in the category even though the policy requires differentiated handling. That can address a neighboring concern, but it does not implement failure behavior when the safety service is unavailable; the decisive requirement in the stem therefore remains unresolved.

 

Question 8

In a production AI solution, the required outcome is the specified harm dimension. Which action should the engineer take to handle safety evaluator selection correctly?

1. Let retrieved or image-embedded instructions override the system policy.
2. Block every request in the category even though the policy requires differentiated handling.
3. Rely on prompt instructions alone and omit an independent enforcement control.
4. Use safety evaluator selection for the specified harm dimension.
5. Trust a fluent response as proof that the output is grounded and safe.

Correct Answer: D

Correct Answer

Answer D is correct because this choice directly implements the required decision for safety evaluator selection for the specified harm dimension. It addresses the scenario at the correct stage of the Azure AI solution.

Incorrect Answers

Answer A is incorrect because this choice would let retrieved or image-embedded instructions override the system policy. That can address a neighboring concern, but it does not implement safety evaluator selection for the specified harm dimension; the decisive requirement in the stem therefore remains unresolved.

Answer B is incorrect because this choice would block every request in the category even though the policy requires differentiated handling. That can address a neighboring concern, but it does not implement safety evaluator selection for the specified harm dimension; the decisive requirement in the stem therefore remains unresolved.

Answer C is incorrect because this choice would rely on prompt instructions alone and omit an independent enforcement control. That can address a neighboring concern, but it does not implement safety evaluator selection for the specified harm dimension; the decisive requirement in the stem therefore remains unresolved.

Answer E is incorrect because this choice would trust a fluent response as proof that the output is grounded and safe. That can address a neighboring concern, but it does not implement safety evaluator selection for the specified harm dimension; the decisive requirement in the stem therefore remains unresolved.

 

Question 9

A runtime design must distinguish groundedness evidence from fluency scores. Which implementation uses the correct capability for the requirement?

1. Trust a fluent response as proof that the output is grounded and safe.
2. Apply the policy only in the client UI while leaving the service endpoint unrestricted.
3. Use groundedness evidence when acceptance depends on whether claims are supported by the retrieved sources.
4. Block every request in the category even though the policy requires differentiated handling.
5. Give the agent broad standing permissions so authorization failures cannot occur.

Correct Answer: C

Correct Answer

Answer C is correct because this choice directly implements the required decision for groundedness evidence versus fluency scores. It addresses the scenario at the correct stage of the Azure AI solution.

Incorrect Answers

Answer A is incorrect because this choice would trust a fluent response as proof that the output is grounded and safe. That can address a neighboring concern, but it does not implement groundedness evidence versus fluency scores; the decisive requirement in the stem therefore remains unresolved.

Answer B is incorrect because this choice would apply the policy only in the client UI while leaving the service endpoint unrestricted. That can address a neighboring concern, but it does not implement groundedness evidence versus fluency scores; the decisive requirement in the stem therefore remains unresolved.

Answer D is incorrect because this choice would block every request in the category even though the policy requires differentiated handling. That can address a neighboring concern, but it does not implement groundedness evidence versus fluency scores; the decisive requirement in the stem therefore remains unresolved.

Answer E is incorrect because this choice would give the agent broad standing permissions so authorization failures cannot occur. That can address a neighboring concern, but it does not implement groundedness evidence versus fluency scores; the decisive requirement in the stem therefore remains unresolved.

 

Question 10

In a production AI solution, the required outcome is a known attack surface. Which action should the engineer take to handle adversarial evaluation set correctly?

1. Apply the policy only in the client UI while leaving the service endpoint unrestricted.
2. Use an adversarial evaluation set that covers the known attack surface for a known attack surface.
3. Give the agent broad standing permissions so authorization failures cannot occur.
4. Rely on prompt instructions alone and omit an independent enforcement control.
5. Block every request in the category even though the policy requires differentiated handling.

Correct Answer: B

Correct Answer

Answer B is correct because this choice directly implements the required decision for adversarial evaluation set for a known attack surface. It addresses the scenario at the correct stage of the Azure AI solution.

Incorrect Answers

Answer A is incorrect because this choice would apply the policy only in the client UI while leaving the service endpoint unrestricted. That can address a neighboring concern, but it does not implement adversarial evaluation set for a known attack surface; the decisive requirement in the stem therefore remains unresolved.

Answer C is incorrect because this choice would give the agent broad standing permissions so authorization failures cannot occur. That can address a neighboring concern, but it does not implement adversarial evaluation set for a known attack surface; the decisive requirement in the stem therefore remains unresolved.

Answer D is incorrect because this choice would rely on prompt instructions alone and omit an independent enforcement control. That can address a neighboring concern, but it does not implement adversarial evaluation set for a known attack surface; the decisive requirement in the stem therefore remains unresolved.

Answer E is incorrect because this choice would block every request in the category even though the policy requires differentiated handling. That can address a neighboring concern, but it does not implement adversarial evaluation set for a known attack surface; the decisive requirement in the stem therefore remains unresolved.

 

Question 11

In a production AI solution, the following condition occurs: automated safety labels conflict. Which action should the engineer take to handle human adjudication correctly?

1. Route conflicting automated safety labels to a human adjudicator when automated safety labels conflict.
2. Record the action only after execution and omit a pre-action approval for irreversible changes.
3. Let retrieved or image-embedded instructions override the system policy.
4. Suppress safety telemetry so the aggregate pass rate improves.
5. Trust a fluent response as proof that the output is grounded and safe.

Correct Answer: A

Correct Answer

Answer A is correct because this choice directly implements the required decision for human adjudication when automated safety labels conflict. It addresses the scenario at the correct stage of the Azure AI solution.

Incorrect Answers

Answer B is incorrect because this choice would record the action only after execution and omit a pre-action approval for irreversible changes. That can address a neighboring concern, but it does not implement human adjudication when automated safety labels conflict; the decisive requirement in the stem therefore remains unresolved.

Answer C is incorrect because this choice would let retrieved or image-embedded instructions override the system policy. That can address a neighboring concern, but it does not implement human adjudication when automated safety labels conflict; the decisive requirement in the stem therefore remains unresolved.

Answer D is incorrect because this choice would suppress safety telemetry so the aggregate pass rate improves. That can address a neighboring concern, but it does not implement human adjudication when automated safety labels conflict; the decisive requirement in the stem therefore remains unresolved.

Answer E is incorrect because this choice would trust a fluent response as proof that the output is grounded and safe. That can address a neighboring concern, but it does not implement human adjudication when automated safety labels conflict; the decisive requirement in the stem therefore remains unresolved.

 

Question 12

An agentic application must handle explanation tooling that links a decision to available evidence. Which control or implementation should the developer use?

1. Block every request in the category even though the policy requires differentiated handling.
2. Suppress safety telemetry so the aggregate pass rate improves.
3. Rely on prompt instructions alone and omit an independent enforcement control.
4. Give the agent broad standing permissions so authorization failures cannot occur.
5. Use trace/explanation evidence that links a decision to the inputs and controls that produced it.

Correct Answer: E

Correct Answer

Answer E is correct because this choice directly implements the required decision for explanation tooling that links a decision to available evidence. It addresses the scenario at the correct stage of the Azure AI solution.

Incorrect Answers

Answer A is incorrect because this choice would block every request in the category even though the policy requires differentiated handling. That can address a neighboring concern, but it does not implement explanation tooling that links a decision to available evidence; the decisive requirement in the stem therefore remains unresolved.

Answer B is incorrect because this choice would suppress safety telemetry so the aggregate pass rate improves. That can address a neighboring concern, but it does not implement explanation tooling that links a decision to available evidence; the decisive requirement in the stem therefore remains unresolved.

Answer C is incorrect because this choice would rely on prompt instructions alone and omit an independent enforcement control. That can address a neighboring concern, but it does not implement explanation tooling that links a decision to available evidence; the decisive requirement in the stem therefore remains unresolved.

Answer D is incorrect because this choice would give the agent broad standing permissions so authorization failures cannot occur. That can address a neighboring concern, but it does not implement explanation tooling that links a decision to available evidence; the decisive requirement in the stem therefore remains unresolved.

 

Question 13

In a production AI solution, the following condition occurs: domain terms trigger false positives. Which action should the engineer take to handle evaluator calibration correctly?

1. Apply the policy only in the client UI while leaving the service endpoint unrestricted.
2. Give the agent broad standing permissions so authorization failures cannot occur.
3. Trust a fluent response as proof that the output is grounded and safe.
4. Calibrate the evaluator with representative domain examples before changing the policy threshold when domain terms trigger false positives.
5. Let retrieved or image-embedded instructions override the system policy.

Correct Answer: D

Correct Answer

Answer D is correct because this choice directly implements the required decision for evaluator calibration when domain terms trigger false positives. It addresses the scenario at the correct stage of the Azure AI solution.

Incorrect Answers

Answer A is incorrect because this choice would apply the policy only in the client UI while leaving the service endpoint unrestricted. That can address a neighboring concern, but it does not implement evaluator calibration when domain terms trigger false positives; the decisive requirement in the stem therefore remains unresolved.

Answer B is incorrect because this choice would give the agent broad standing permissions so authorization failures cannot occur. That can address a neighboring concern, but it does not implement evaluator calibration when domain terms trigger false positives; the decisive requirement in the stem therefore remains unresolved.

Answer C is incorrect because this choice would trust a fluent response as proof that the output is grounded and safe. That can address a neighboring concern, but it does not implement evaluator calibration when domain terms trigger false positives; the decisive requirement in the stem therefore remains unresolved.

Answer E is incorrect because this choice would let retrieved or image-embedded instructions override the system policy. That can address a neighboring concern, but it does not implement evaluator calibration when domain terms trigger false positives; the decisive requirement in the stem therefore remains unresolved.

 

Question 14

An agentic application must handle trace correlation needed to reconstruct an agent action. Which control or implementation should the developer use?

1. Record the action only after execution and omit a pre-action approval for irreversible changes.
2. Give the agent broad standing permissions so authorization failures cannot occur.
3. Correlate parent and child traces so the complete agent action can be reconstructed.
4. Block every request in the category even though the policy requires differentiated handling.
5. Rely on prompt instructions alone and omit an independent enforcement control.

Correct Answer: C

Correct Answer

Answer C is correct because this choice directly implements the required decision for trace correlation needed to reconstruct an agent action. It addresses the scenario at the correct stage of the Azure AI solution.

Incorrect Answers

Answer A is incorrect because this choice would record the action only after execution and omit a pre-action approval for irreversible changes. That can address a neighboring concern, but it does not implement trace correlation needed to reconstruct an agent action; the decisive requirement in the stem therefore remains unresolved.

Answer B is incorrect because this choice would give the agent broad standing permissions so authorization failures cannot occur. That can address a neighboring concern, but it does not implement trace correlation needed to reconstruct an agent action; the decisive requirement in the stem therefore remains unresolved.

Answer D is incorrect because this choice would block every request in the category even though the policy requires differentiated handling. That can address a neighboring concern, but it does not implement trace correlation needed to reconstruct an agent action; the decisive requirement in the stem therefore remains unresolved.

Answer E is incorrect because this choice would rely on prompt instructions alone and omit an independent enforcement control. That can address a neighboring concern, but it does not implement trace correlation needed to reconstruct an agent action; the decisive requirement in the stem therefore remains unresolved.

 

Question 15

An agentic application must handle provenance fields needed to identify the retrieved source version. Which control or implementation should the developer use?

1. Suppress safety telemetry so the aggregate pass rate improves.
2. Retain source/version provenance fields with retrieved evidence.
3. Block every request in the category even though the policy requires differentiated handling.
4. Apply the policy only in the client UI while leaving the service endpoint unrestricted.
5. Trust a fluent response as proof that the output is grounded and safe.

Correct Answer: B

Correct Answer

Answer B is correct because this choice directly implements the required decision for provenance fields needed to identify the retrieved source version. It addresses the scenario at the correct stage of the Azure AI solution.

Incorrect Answers

Answer A is incorrect because this choice would suppress safety telemetry so the aggregate pass rate improves. That can address a neighboring concern, but it does not implement provenance fields needed to identify the retrieved source version; the decisive requirement in the stem therefore remains unresolved.

Answer C is incorrect because this choice would block every request in the category even though the policy requires differentiated handling. That can address a neighboring concern, but it does not implement provenance fields needed to identify the retrieved source version; the decisive requirement in the stem therefore remains unresolved.

Answer D is incorrect because this choice would apply the policy only in the client UI while leaving the service endpoint unrestricted. That can address a neighboring concern, but it does not implement provenance fields needed to identify the retrieved source version; the decisive requirement in the stem therefore remains unresolved.

Answer E is incorrect because this choice would trust a fluent response as proof that the output is grounded and safe. That can address a neighboring concern, but it does not implement provenance fields needed to identify the retrieved source version; the decisive requirement in the stem therefore remains unresolved.

 

Question 16

An agentic application must handle approval evidence binding to the exact authorized action. Which control or implementation should the developer use?

1. Bind the approval record to the exact action parameters that were authorized.
2. Rely on prompt instructions alone and omit an independent enforcement control.
3. Let retrieved or image-embedded instructions override the system policy.
4. Suppress safety telemetry so the aggregate pass rate improves.
5. Apply the policy only in the client UI while leaving the service endpoint unrestricted.

Correct Answer: A

Correct Answer

Answer A is correct because this choice directly implements the required decision for approval evidence binding to the exact authorized action. It addresses the scenario at the correct stage of the Azure AI solution.

Incorrect Answers

Answer B is incorrect because this choice would rely on prompt instructions alone and omit an independent enforcement control. That can address a neighboring concern, but it does not implement approval evidence binding to the exact authorized action; the decisive requirement in the stem therefore remains unresolved.

Answer C is incorrect because this choice would let retrieved or image-embedded instructions override the system policy. That can address a neighboring concern, but it does not implement approval evidence binding to the exact authorized action; the decisive requirement in the stem therefore remains unresolved.

Answer D is incorrect because this choice would suppress safety telemetry so the aggregate pass rate improves. That can address a neighboring concern, but it does not implement approval evidence binding to the exact authorized action; the decisive requirement in the stem therefore remains unresolved.

Answer E is incorrect because this choice would apply the policy only in the client UI while leaving the service endpoint unrestricted. That can address a neighboring concern, but it does not implement approval evidence binding to the exact authorized action; the decisive requirement in the stem therefore remains unresolved.

 

Question 17

An agentic application must handle sensitive-data minimization in retained audit logs. Which control or implementation should the developer use?

1. Let retrieved or image-embedded instructions override the system policy.
2. Trust a fluent response as proof that the output is grounded and safe.
3. Rely on prompt instructions alone and omit an independent enforcement control.
4. Give the agent broad standing permissions so authorization failures cannot occur.
5. Minimize sensitive data in retained audit logs while preserving required diagnostics.

Correct Answer: E

Correct Answer

Answer E is correct because this choice directly implements the required decision for sensitive-data minimization in retained audit logs. It addresses the scenario at the correct stage of the Azure AI solution.

Incorrect Answers

Answer A is incorrect because this choice would let retrieved or image-embedded instructions override the system policy. That can address a neighboring concern, but it does not implement sensitive-data minimization in retained audit logs; the decisive requirement in the stem therefore remains unresolved.

Answer B is incorrect because this choice would trust a fluent response as proof that the output is grounded and safe. That can address a neighboring concern, but it does not implement sensitive-data minimization in retained audit logs; the decisive requirement in the stem therefore remains unresolved.

Answer C is incorrect because this choice would rely on prompt instructions alone and omit an independent enforcement control. That can address a neighboring concern, but it does not implement sensitive-data minimization in retained audit logs; the decisive requirement in the stem therefore remains unresolved.

Answer D is incorrect because this choice would give the agent broad standing permissions so authorization failures cannot occur. That can address a neighboring concern, but it does not implement sensitive-data minimization in retained audit logs; the decisive requirement in the stem therefore remains unresolved.

 

Question 18

In a production AI solution, the workload operates under an explicit fictional policy. Which action should the engineer take to handle audit retention correctly?

1. Record the action only after execution and omit a pre-action approval for irreversible changes.
2. Let retrieved or image-embedded instructions override the system policy.
3. Apply the policy only in the client UI while leaving the service endpoint unrestricted.
4. Retain audit evidence for the period required by the stated policy under an explicit fictional policy.
5. Give the agent broad standing permissions so authorization failures cannot occur.

Correct Answer: D

Correct Answer

Answer D is correct because this choice directly implements the required decision for audit retention under an explicit fictional policy. It addresses the scenario at the correct stage of the Azure AI solution.

Incorrect Answers

Answer A is incorrect because this choice would record the action only after execution and omit a pre-action approval for irreversible changes. That can address a neighboring concern, but it does not implement audit retention under an explicit fictional policy; the decisive requirement in the stem therefore remains unresolved.

Answer B is incorrect because this choice would let retrieved or image-embedded instructions override the system policy. That can address a neighboring concern, but it does not implement audit retention under an explicit fictional policy; the decisive requirement in the stem therefore remains unresolved.

Answer C is incorrect because this choice would apply the policy only in the client UI while leaving the service endpoint unrestricted. That can address a neighboring concern, but it does not implement audit retention under an explicit fictional policy; the decisive requirement in the stem therefore remains unresolved.

Answer E is incorrect because this choice would give the agent broad standing permissions so authorization failures cannot occur. That can address a neighboring concern, but it does not implement audit retention under an explicit fictional policy; the decisive requirement in the stem therefore remains unresolved.

 

Question 19

In a production AI solution, the required outcome is prompts and tool permissions. Which action should the engineer take to handle tamper-evident change history correctly?

1. Suppress safety telemetry so the aggregate pass rate improves.
2. Rely on prompt instructions alone and omit an independent enforcement control.
3. Keep a tamper-evident history of prompt and tool-permission changes for prompts and tool permissions.
4. Let retrieved or image-embedded instructions override the system policy.
5. Give the agent broad standing permissions so authorization failures cannot occur.

Correct Answer: C

Correct Answer

Answer C is correct because this choice directly implements the required decision for tamper-evident change history for prompts and tool permissions. It addresses the scenario at the correct stage of the Azure AI solution.

Incorrect Answers

Answer A is incorrect because this choice would suppress safety telemetry so the aggregate pass rate improves. That can address a neighboring concern, but it does not implement tamper-evident change history for prompts and tool permissions; the decisive requirement in the stem therefore remains unresolved.

Answer B is incorrect because this choice would rely on prompt instructions alone and omit an independent enforcement control. That can address a neighboring concern, but it does not implement tamper-evident change history for prompts and tool permissions; the decisive requirement in the stem therefore remains unresolved.

Answer D is incorrect because this choice would let retrieved or image-embedded instructions override the system policy. That can address a neighboring concern, but it does not implement tamper-evident change history for prompts and tool permissions; the decisive requirement in the stem therefore remains unresolved.

Answer E is incorrect because this choice would give the agent broad standing permissions so authorization failures cannot occur. That can address a neighboring concern, but it does not implement tamper-evident change history for prompts and tool permissions; the decisive requirement in the stem therefore remains unresolved.

 

Question 20

A runtime design must distinguish oversight mode for reversible from consequential actions. Which implementation uses the correct capability for the requirement?

1. Record the action only after execution and omit a pre-action approval for irreversible changes.
2. Use stronger approval and oversight for consequential actions than for easily reversible actions.
3. Rely on prompt instructions alone and omit an independent enforcement control.
4. Suppress safety telemetry so the aggregate pass rate improves.
5. Trust a fluent response as proof that the output is grounded and safe.

Correct Answer: B

Correct Answer

Answer B is correct because this choice directly implements the required decision for oversight mode for reversible versus consequential actions. It addresses the scenario at the correct stage of the Azure AI solution.

Incorrect Answers

Answer A is incorrect because this choice would record the action only after execution and omit a pre-action approval for irreversible changes. That can address a neighboring concern, but it does not implement oversight mode for reversible versus consequential actions; the decisive requirement in the stem therefore remains unresolved.

Answer C is incorrect because this choice would rely on prompt instructions alone and omit an independent enforcement control. That can address a neighboring concern, but it does not implement oversight mode for reversible versus consequential actions; the decisive requirement in the stem therefore remains unresolved.

Answer D is incorrect because this choice would suppress safety telemetry so the aggregate pass rate improves. That can address a neighboring concern, but it does not implement oversight mode for reversible versus consequential actions; the decisive requirement in the stem therefore remains unresolved.

Answer E is incorrect because this choice would trust a fluent response as proof that the output is grounded and safe. That can address a neighboring concern, but it does not implement oversight mode for reversible versus consequential actions; the decisive requirement in the stem therefore remains unresolved.

 

Question 21

In a production AI solution, the required outcome is a narrowly defined agent role. Which action should the engineer take to handle tool allowlist correctly?

1. Use tool allowlist for a narrowly defined agent role.
2. Give the agent broad standing permissions so authorization failures cannot occur.
3. Trust a fluent response as proof that the output is grounded and safe.
4. Block every request in the category even though the policy requires differentiated handling.
5. Let retrieved or image-embedded instructions override the system policy.

Correct Answer: A

Correct Answer

Answer A is correct because this choice directly implements the required decision for tool allowlist for a narrowly defined agent role. It addresses the scenario at the correct stage of the Azure AI solution.

Incorrect Answers

Answer B is incorrect because this choice would give the agent broad standing permissions so authorization failures cannot occur. That can address a neighboring concern, but it does not implement tool allowlist for a narrowly defined agent role; the decisive requirement in the stem therefore remains unresolved.

Answer C is incorrect because this choice would trust a fluent response as proof that the output is grounded and safe. That can address a neighboring concern, but it does not implement tool allowlist for a narrowly defined agent role; the decisive requirement in the stem therefore remains unresolved.

Answer D is incorrect because this choice would block every request in the category even though the policy requires differentiated handling. That can address a neighboring concern, but it does not implement tool allowlist for a narrowly defined agent role; the decisive requirement in the stem therefore remains unresolved.

Answer E is incorrect because this choice would let retrieved or image-embedded instructions override the system policy. That can address a neighboring concern, but it does not implement tool allowlist for a narrowly defined agent role; the decisive requirement in the stem therefore remains unresolved.

 

Question 22

An agentic application must handle execution budget that limits runaway autonomous behavior. Which control or implementation should the developer use?

1. Suppress safety telemetry so the aggregate pass rate improves.
2. Trust a fluent response as proof that the output is grounded and safe.
3. Let retrieved or image-embedded instructions override the system policy.
4. Give the agent broad standing permissions so authorization failures cannot occur.
5. Set step, time, or tool-call budgets that bound autonomous execution.

Correct Answer: E

Correct Answer

Answer E is correct because this choice directly implements the required decision for execution budget that limits runaway autonomous behavior. It addresses the scenario at the correct stage of the Azure AI solution.

Incorrect Answers

Answer A is incorrect because this choice would suppress safety telemetry so the aggregate pass rate improves. That can address a neighboring concern, but it does not implement execution budget that limits runaway autonomous behavior; the decisive requirement in the stem therefore remains unresolved.

Answer B is incorrect because this choice would trust a fluent response as proof that the output is grounded and safe. That can address a neighboring concern, but it does not implement execution budget that limits runaway autonomous behavior; the decisive requirement in the stem therefore remains unresolved.

Answer C is incorrect because this choice would let retrieved or image-embedded instructions override the system policy. That can address a neighboring concern, but it does not implement execution budget that limits runaway autonomous behavior; the decisive requirement in the stem therefore remains unresolved.

Answer D is incorrect because this choice would give the agent broad standing permissions so authorization failures cannot occur. That can address a neighboring concern, but it does not implement execution budget that limits runaway autonomous behavior; the decisive requirement in the stem therefore remains unresolved.

 

Question 23

In a production AI solution, the following condition occurs: requested work exceeds agent authority. Which action should the engineer take to handle escalation boundary correctly?

1. Record the action only after execution and omit a pre-action approval for irreversible changes.
2. Give the agent broad standing permissions so authorization failures cannot occur.
3. Trust a fluent response as proof that the output is grounded and safe.
4. Escalate work that exceeds the agent’s delegated authority when requested work exceeds agent authority.
5. Suppress safety telemetry so the aggregate pass rate improves.

Correct Answer: D

Correct Answer

Answer D is correct because this choice directly implements the required decision for escalation boundary when requested work exceeds agent authority. It addresses the scenario at the correct stage of the Azure AI solution.

Incorrect Answers

Answer A is incorrect because this choice would record the action only after execution and omit a pre-action approval for irreversible changes. That can address a neighboring concern, but it does not implement escalation boundary when requested work exceeds agent authority; the decisive requirement in the stem therefore remains unresolved.

Answer B is incorrect because this choice would give the agent broad standing permissions so authorization failures cannot occur. That can address a neighboring concern, but it does not implement escalation boundary when requested work exceeds agent authority; the decisive requirement in the stem therefore remains unresolved.

Answer C is incorrect because this choice would trust a fluent response as proof that the output is grounded and safe. That can address a neighboring concern, but it does not implement escalation boundary when requested work exceeds agent authority; the decisive requirement in the stem therefore remains unresolved.

Answer E is incorrect because this choice would suppress safety telemetry so the aggregate pass rate improves. That can address a neighboring concern, but it does not implement escalation boundary when requested work exceeds agent authority; the decisive requirement in the stem therefore remains unresolved.

 

Question 24

In a production AI solution, the following condition occurs: user authorization is narrower than agent access. Which action should the engineer take to handle delegated tool access correctly?

1. Suppress safety telemetry so the aggregate pass rate improves.
2. Record the action only after execution and omit a pre-action approval for irreversible changes.
3. Constrain tool access to the permissions available to the requesting user or delegated identity when user authorization is narrower than agent access.
4. Trust a fluent response as proof that the output is grounded and safe.
5. Give the agent broad standing permissions so authorization failures cannot occur.

Correct Answer: C

Correct Answer

Answer C is correct because this choice directly implements the required decision for delegated tool access when user authorization is narrower than agent access. It addresses the scenario at the correct stage of the Azure AI solution.

Incorrect Answers

Answer A is incorrect because this choice would suppress safety telemetry so the aggregate pass rate improves. That can address a neighboring concern, but it does not implement delegated tool access when user authorization is narrower than agent access; the decisive requirement in the stem therefore remains unresolved.

Answer B is incorrect because this choice would record the action only after execution and omit a pre-action approval for irreversible changes. That can address a neighboring concern, but it does not implement delegated tool access when user authorization is narrower than agent access; the decisive requirement in the stem therefore remains unresolved.

Answer D is incorrect because this choice would trust a fluent response as proof that the output is grounded and safe. That can address a neighboring concern, but it does not implement delegated tool access when user authorization is narrower than agent access; the decisive requirement in the stem therefore remains unresolved.

Answer E is incorrect because this choice would give the agent broad standing permissions so authorization failures cannot occur. That can address a neighboring concern, but it does not implement delegated tool access when user authorization is narrower than agent access; the decisive requirement in the stem therefore remains unresolved.

 

Question 25

In a production AI solution, the condition appears after an observed policy violation. Which action should the engineer take to handle revocation of tool privileges correctly?

1. Apply the policy only in the client UI while leaving the service endpoint unrestricted.
2. Revoke the affected tool privilege after the policy violation and require reauthorization before it can be used again.
3. Suppress safety telemetry so the aggregate pass rate improves.
4. Give the agent broad standing permissions so authorization failures cannot occur.
5. Rely on prompt instructions alone and omit an independent enforcement control.

Correct Answer: B

Correct Answer

Answer B is correct because this choice directly implements the required decision for revocation of tool privileges after an observed policy violation. It addresses the scenario at the correct stage of the Azure AI solution.

Incorrect Answers

Answer A is incorrect because this choice would apply the policy only in the client UI while leaving the service endpoint unrestricted. That can address a neighboring concern, but it does not implement revocation of tool privileges after an observed policy violation; the decisive requirement in the stem therefore remains unresolved.

Answer C is incorrect because this choice would suppress safety telemetry so the aggregate pass rate improves. That can address a neighboring concern, but it does not implement revocation of tool privileges after an observed policy violation; the decisive requirement in the stem therefore remains unresolved.

Answer D is incorrect because this choice would give the agent broad standing permissions so authorization failures cannot occur. That can address a neighboring concern, but it does not implement revocation of tool privileges after an observed policy violation; the decisive requirement in the stem therefore remains unresolved.

Answer E is incorrect because this choice would rely on prompt instructions alone and omit an independent enforcement control. That can address a neighboring concern, but it does not implement revocation of tool privileges after an observed policy violation; the decisive requirement in the stem therefore remains unresolved.