Practice Exams:

Microsoft AI-103 Dumps

EXAM: Developing AI Apps and Agents on Azure
Verified By Experts
Premium Bundle
$49.99

AI-103 Premium ETE File


  • Premium File 187 Questions & Answers
Whats Included:
  • Latest Questions
  • 100% Accurate Answers
  • Fast Exam Updates

Microsoft AI-103 Practice Test 20

Question 190

During a architecture migration, a security engineer must accept a change for data-connector permissions. The issue to be addressed is: An agent can read documents from projects it was never assigned. Which concrete verification action demonstrates that the relevant behavior works?

  1. Reuse a tenant-wide reader role for every agent knowledge source
  2. Ask an agent to retrieve forbidden document IDs and confirm access is blocked
  3. Restrict retrieval connector data-plane permissions to authorized corpora
  4. Inspect a slow trace and identify the tool step consuming the delay

Explanation

Correct Answer

B:

To verify this change, the insurance platform should ask an agent to retrieve forbidden document IDs and confirm access is blocked. Operators still need the implementation to restrict retrieval connector data-plane permissions to authorized corpora and the preventive requirement to test cross-project denial and log knowledge-source permission changes.

Incorrect Answers

A:

In the insurance platform acceptance review, trying to reuse a tenant-wide reader role for every agent knowledge source does not test the exact data-connector permissions condition where an agent can read documents from projects it was never assigned; the relevant test is to ask an agent to retrieve forbidden document IDs and confirm access is blocked. Acceptance should instead demonstrate that the service can ask an agent to retrieve forbidden document IDs and confirm access is blocked, using the described failure as the test criterion.

C:

The action to restrict retrieval connector data-plane permissions to authorized corpora can contribute to data-connector permissions as a implementation step; here, an agent can read documents from projects it was never assigned. In the insurance platform acceptance review, trying to restrict retrieval connector data-plane permissions to authorized corpora does not test the exact data-connector permissions condition where an agent can read documents from projects it was never assigned; the relevant test is to ask an agent to retrieve forbidden document IDs and confirm access is blocked.

D:

For agent tool latency, it can be sensible to inspect a slow trace and identify the tool step consuming the delay if an otherwise fast chatbot responds slowly after retrieving external data. Without the actual change to ask an agent to retrieve forbidden document IDs and confirm access is blocked, the answer lacks the correct operational scope even if the alternate action is technically valid elsewhere.

Question 191

At a research university, the agent designer is preparing a source update. The concrete requirement is: A retrieved page contains instructions attempting to override an agent system policy. Which design action BEST addresses the stated need?

  1. Treat retrieved text as untrusted content and evaluate prompt-injection resistance
  2. Test adversarial passages that request secret disclosure or permission escalation
  3. Combine functional synthetic inference checks with service error and saturation metrics
  4. Execute instructions from a retrieved page as though they were developer commands

Explanation

Correct Answer

A:

The proposed implementation is to treat retrieved text as untrusted content and evaluate prompt-injection resistance; this matches the operational requirement that a retrieved page contains instructions attempting to override an agent system policy. The team can test adversarial passages that request secret disclosure or permission escalation. The team should also evaluate retrieval-time injection scenarios and require explicit tool allowlists.

Incorrect Answers

B:

The action to test adversarial passages that request secret disclosure or permission escalation can contribute to grounding attack monitoring as a targeted acceptance check step; here, a retrieved page contains instructions attempting to override an agent system policy. Choosing to test adversarial passages that request secret disclosure or permission escalation cannot directly replace the specified action to treat retrieved text as untrusted content and evaluate prompt-injection resistance when a retrieved page contains instructions attempting to override an agent system policy.

C:

Choosing to combine functional synthetic inference checks with service error and saturation metrics would direct effort toward deployment availability, whose characteristic issue is that a regional service has intermittent errors but its health endpoint reports green. The relevant corrective or preventive action here remains to treat retrieved text as untrusted content and evaluate prompt-injection resistance, with its own validation criteria.

D:

Choosing to execute instructions from a retrieved page as though they were developer commands cannot directly replace the specified action to treat retrieved text as untrusted content and evaluate prompt-injection resistance when a retrieved page contains instructions attempting to override an agent system policy. The actual implementation must treat retrieved text as untrusted content and evaluate prompt-injection resistance, because the stated service condition is that a retrieved page contains instructions attempting to override an agent system policy.

Question 192

After a cost review at a education technology company, the team observes this failure: A retrieved page contains instructions attempting to override an agent system policy. Which flawed assumption or implementation choice would MOST DIRECTLY explain why the intended outcome was not achieved?

  1. Treat every user timeout as identical model endpoint unavailability
  2. Execute instructions from a retrieved page as though they were developer commands
  3. Test adversarial passages that request secret disclosure or permission escalation
  4. Treat retrieved text as untrusted content and evaluate prompt-injection resistance

Explanation

Correct Answer

B:

This design mistake is to execute instructions from a retrieved page as though they were developer commands; it conflicts with the actual grounding attack monitoring behavior needed when a retrieved page contains instructions attempting to override an agent system policy. The education technology company should instead treat retrieved text as untrusted content and evaluate prompt-injection resistance, followed by a check to test adversarial passages that request secret disclosure or permission escalation.

Incorrect Answers

A:

The option to treat every user timeout as identical model endpoint unavailability would normally serve a different provider health monitoring requirement, where an app needs to distinguish model failures from search and speech outages. The alternative may be useful in its own context, but it does not directly supply execute instructions from a retrieved page as though they were developer commands for the stated decision.

C:

The alternative to test adversarial passages that request secret disclosure or permission escalation is a targeted acceptance check step rather than the faulty assumption being diagnosed. This education technology company failure-analysis alternative of test adversarial passages that request secret disclosure or permission escalation does not diagnose the particular grounding attack monitoring defect where a retrieved page contains instructions attempting to override an agent system policy; the keyed mistake is execute instructions from a retrieved page as though they were developer commands.

D:

The alternative to treat retrieved text as untrusted content and evaluate prompt-injection resistance is a implementation step rather than the faulty assumption being diagnosed. This education technology company failure-analysis alternative of treat retrieved text as untrusted content and evaluate prompt-injection resistance does not diagnose the particular grounding attack monitoring defect where a retrieved page contains instructions attempting to override an agent system policy; the keyed mistake is execute instructions from a retrieved page as though they were developer commands.

Question 193

The pharmaceutical research group is responding to a reliability exercise. A solution architect discovers this issue: A retrieved page contains instructions attempting to override an agent system policy. What corrective step should the team take to restore the intended functionality most directly?

  1. Fix deployed identity bindings and regenerate correctly scoped access tokens
  2. Strip dangerous source instructions and reinforce grounded tool-use boundaries
  3. Test adversarial passages that request secret disclosure or permission escalation
  4. Execute instructions from a retrieved page as though they were developer commands

Explanation

Correct Answer

B:

The direct repair is to strip dangerous source instructions and reinforce grounded tool-use boundaries. Once the condition is corrected, the pharmaceutical research group should test adversarial passages that request secret disclosure or permission escalation. To reduce recurrence, it should also evaluate retrieval-time injection scenarios and require explicit tool allowlists; those steps provide different recovery and prevention functions.

Incorrect Answers

A:

This approach relates to access token authorization: an inference app succeeds locally using developer credentials but fails under deployment. If the team chose it instead, the targeted check to test adversarial passages that request secret disclosure or permission escalation could still fail because the underlying required action is to strip dangerous source instructions and reinforce grounded tool-use boundaries.

C:

The action to test adversarial passages that request secret disclosure or permission escalation can contribute to grounding attack monitoring as a targeted acceptance check step; here, a retrieved page contains instructions attempting to override an agent system policy. During pharmaceutical research group remediation, the proposed step to test adversarial passages that request secret disclosure or permission escalation is not the direct restoration action for grounding attack monitoring when a retrieved page contains instructions attempting to override an agent system policy; the keyed repair is strip dangerous source instructions and reinforce grounded tool-use boundaries.

D:

During pharmaceutical research group remediation, the proposed step to execute instructions from a retrieved page as though they were developer commands is not the direct restoration action for grounding attack monitoring when a retrieved page contains instructions attempting to override an agent system policy; the keyed repair is strip dangerous source instructions and reinforce grounded tool-use boundaries. After remediation the team must test adversarial passages that request secret disclosure or permission escalation; the action in this option does not itself establish that target recovery outcome.

Question 194

A machine-learning developer at a travel support provider wants the following issue to be prevented in later deployments: A retrieved page contains instructions attempting to override an agent system policy. Which enduring control or engineering guardrail BEST prevents recurrence?

  1. Strip dangerous source instructions and reinforce grounded tool-use boundaries
  2. Execute instructions from a retrieved page as though they were developer commands
  3. Evaluate retrieval-time injection scenarios and require explicit tool allowlists
  4. Require a stable regression test suite and deployment annotations

Explanation

Correct Answer

C:

Standardizing the boundary to evaluate retrieval-time injection scenarios and require explicit tool allowlists is preventive because the same grounding attack monitoring failure could otherwise return: a retrieved page contains instructions attempting to override an agent system policy. For this travel support provider, the service should also treat retrieved text as untrusted content and evaluate prompt-injection resistance, and tests should test adversarial passages that request secret disclosure or permission escalation so that the ongoing requirement is measurable.

Incorrect Answers

A:

The action to strip dangerous source instructions and reinforce grounded tool-use boundaries can contribute to grounding attack monitoring as a corrective remediation step; here, a retrieved page contains instructions attempting to override an agent system policy. For prevention at the travel support provider, the measure to strip dangerous source instructions and reinforce grounded tool-use boundaries is weaker or aimed at another issue than the grounding attack monitoring guardrail to evaluate retrieval-time injection scenarios and require explicit tool allowlists needed when a retrieved page contains instructions attempting to override an agent system policy.

B:

For prevention at the travel support provider, the measure to execute instructions from a retrieved page as though they were developer commands is weaker or aimed at another issue than the grounding attack monitoring guardrail to evaluate retrieval-time injection scenarios and require explicit tool allowlists needed when a retrieved page contains instructions attempting to override an agent system policy. The lasting guardrail for this specific threat is to evaluate retrieval-time injection scenarios and require explicit tool allowlists, not simply to complete an adjacent service task.

D:

The action to require a stable regression test suite and deployment annotations addresses model output drift rather than the decisive grounding attack monitoring condition. The travel support provider should not infer that fixing one adjacent AI service validates its unrelated workflow; the requested action is to evaluate retrieval-time injection scenarios and require explicit tool allowlists.

Question 195

During a cost review, a AI application owner must accept a change for grounding attack monitoring. The issue to be addressed is: A retrieved page contains instructions attempting to override an agent system policy. Which concrete verification action demonstrates that the relevant behavior works?

  1. Treat retrieved text as untrusted content and evaluate prompt-injection resistance
  2. Compare cited chunks to a labeled answer-evidence evaluation set
  3. Test adversarial passages that request secret disclosure or permission escalation
  4. Execute instructions from a retrieved page as though they were developer commands

Explanation

Correct Answer

C:

The targeted acceptance test is to test adversarial passages that request secret disclosure or permission escalation. The design should first treat retrieved text as untrusted content and evaluate prompt-injection resistance. A robust scientific publisher deployment also needs to evaluate retrieval-time injection scenarios and require explicit tool allowlists; passing this check provides relevant evidence but does not certify every possible input or security risk.

Incorrect Answers

A:

The action to treat retrieved text as untrusted content and evaluate prompt-injection resistance can contribute to grounding attack monitoring as a implementation step; here, a retrieved page contains instructions attempting to override an agent system policy. In the scientific publisher acceptance review, trying to treat retrieved text as untrusted content and evaluate prompt-injection resistance does not test the exact grounding attack monitoring condition where a retrieved page contains instructions attempting to override an agent system policy; the relevant test is to test adversarial passages that request secret disclosure or permission escalation.

B:

For grounding quality monitoring, it can be sensible to compare cited chunks to a labeled answer-evidence evaluation set if a RAG bot returns fluent answers that cite irrelevant source passages. Without the actual change to test adversarial passages that request secret disclosure or permission escalation, the answer lacks the correct operational scope even if the alternate action is technically valid elsewhere.

D:

In the scientific publisher acceptance review, trying to execute instructions from a retrieved page as though they were developer commands does not test the exact grounding attack monitoring condition where a retrieved page contains instructions attempting to override an agent system policy; the relevant test is to test adversarial passages that request secret disclosure or permission escalation. Acceptance should instead demonstrate that the service can test adversarial passages that request secret disclosure or permission escalation, using the described failure as the test criterion.

Question 196

At a insurance platform, the quality assurance lead is preparing a support escalation. The concrete requirement is: An app needs to distinguish model failures from search and speech outages. Which design action BEST addresses the stated need?

  1. Simulate one service failure and verify the correct alert identifies it
  2. Treat every user timeout as identical model endpoint unavailability
  3. Version prompt templates model IDs datasets and evaluation results in each release
  4. Instrument each dependent Foundry and tool service with distinct success-rate metrics

Explanation

Correct Answer

D:

Choosing to instrument each dependent Foundry and tool service with distinct success-rate metrics directly targets the described provider health monitoring gap: an app needs to distinguish model failures from search and speech outages. As a lasting boundary, they also need to define service-specific health SLOs and dependency diagrams, because an implementation by itself does not prove safe operation.

Incorrect Answers

A:

The action to simulate one service failure and verify the correct alert identifies it can contribute to provider health monitoring as a targeted acceptance check step; here, an app needs to distinguish model failures from search and speech outages. Choosing to simulate one service failure and verify the correct alert identifies it cannot directly replace the specified action to instrument each dependent Foundry and tool service with distinct success-rate metrics when an app needs to distinguish model failures from search and speech outages.

B:

Choosing to treat every user timeout as identical model endpoint unavailability cannot directly replace the specified action to instrument each dependent Foundry and tool service with distinct success-rate metrics when an app needs to distinguish model failures from search and speech outages. The actual implementation must instrument each dependent Foundry and tool service with distinct success-rate metrics, because the stated service condition is that an app needs to distinguish model failures from search and speech outages.

C:

Choosing to version prompt templates model IDs datasets and evaluation results in each release would direct effort toward experiment reproducibility, whose characteristic issue is that a team cannot explain which prompt version produced a degraded answer. The relevant corrective or preventive action here remains to instrument each dependent Foundry and tool service with distinct success-rate metrics, with its own validation criteria.

Question 197

After a service audit at a media platform, the team observes this failure: An app needs to distinguish model failures from search and speech outages. Which flawed assumption or implementation choice would MOST DIRECTLY explain why the intended outcome was not achieved?

  1. Simulate one service failure and verify the correct alert identifies it
  2. Instrument each dependent Foundry and tool service with distinct success-rate metrics
  3. Treat every user timeout as identical model endpoint unavailability
  4. Enable full prompt-body logging in a broadly readable workspace by default

Explanation

Correct Answer

C:

The faulty assumption is treat every user timeout as identical model endpoint unavailability. A functioning alternative would be to instrument each dependent Foundry and tool service with distinct success-rate metrics, and the relevant validation would be to simulate one service failure and verify the correct alert identifies it.

Incorrect Answers

A:

The alternative to simulate one service failure and verify the correct alert identifies it is a targeted acceptance check step rather than the faulty assumption being diagnosed. This media platform failure-analysis alternative of simulate one service failure and verify the correct alert identifies it does not diagnose the particular provider health monitoring defect where an app needs to distinguish model failures from search and speech outages; the keyed mistake is treat every user timeout as identical model endpoint unavailability.

B:

The alternative to instrument each dependent Foundry and tool service with distinct success-rate metrics is a implementation step rather than the faulty assumption being diagnosed. This media platform failure-analysis alternative of instrument each dependent Foundry and tool service with distinct success-rate metrics does not diagnose the particular provider health monitoring defect where an app needs to distinguish model failures from search and speech outages; the keyed mistake is treat every user timeout as identical model endpoint unavailability.

D:

This media platform failure-analysis alternative of enable full prompt-body logging in a broadly readable workspace by default does not diagnose the particular provider health monitoring defect where an app needs to distinguish model failures from search and speech outages; the keyed mistake is treat every user timeout as identical model endpoint unavailability. The alternative may be useful in its own context, but it does not directly supply treat every user timeout as identical model endpoint unavailability for the stated decision.

Question 198

The customer support center is responding to a security tabletop exercise. A compliance specialist discovers this issue: An app needs to distinguish model failures from search and speech outages. What corrective step should the team take to restore the intended functionality most directly?

  1. Route incidents to the failing dependency based on correlated traces
  2. Simulate one service failure and verify the correct alert identifies it
  3. Define service-specific health SLOs and dependency diagrams
  4. Treat every user timeout as identical model endpoint unavailability
  5. Optimize the slow tool or apply bounded timeouts and safe fallbacks

Explanation

Correct Answer

A:

The appropriate remediation is to route incidents to the failing dependency based on correlated traces, because the failure concerns provider health monitoring: an app needs to distinguish model failures from search and speech outages. A meaningful verification should simulate one service failure and verify the correct alert identifies it, and the durable operating rule should define service-specific health SLOs and dependency diagrams after the immediate fix.

Incorrect Answers

B:

The action to simulate one service failure and verify the correct alert identifies it can contribute to provider health monitoring as a targeted acceptance check step; here, an app needs to distinguish model failures from search and speech outages. During customer support center remediation, the proposed step to simulate one service failure and verify the correct alert identifies it is not the direct restoration action for provider health monitoring when an app needs to distinguish model failures from search and speech outages; the keyed repair is route incidents to the failing dependency based on correlated traces.

C:

The action to define service-specific health SLOs and dependency diagrams can contribute to provider health monitoring as a ongoing preventive rule step; here, an app needs to distinguish model failures from search and speech outages. During customer support center remediation, the proposed step to define service-specific health SLOs and dependency diagrams is not the direct restoration action for provider health monitoring when an app needs to distinguish model failures from search and speech outages; the keyed repair is route incidents to the failing dependency based on correlated traces.

D:

During customer support center remediation, the proposed step to treat every user timeout as identical model endpoint unavailability is not the direct restoration action for provider health monitoring when an app needs to distinguish model failures from search and speech outages; the keyed repair is route incidents to the failing dependency based on correlated traces. After remediation the team must simulate one service failure and verify the correct alert identifies it; the action in this option does not itself establish that target recovery outcome.

E:

The action to optimize the slow tool or apply bounded timeouts and safe fallbacks addresses agent tool latency rather than the decisive provider health monitoring condition. The customer support center should not infer that fixing one adjacent AI service validates its unrelated workflow; the requested action is to route incidents to the failing dependency based on correlated traces.

Question 199

A product support lead at a manufacturing group wants the following issue to be prevented in later deployments: An app needs to distinguish model failures from search and speech outages. Which enduring control or engineering guardrail BEST prevents recurrence?

  1. Establish token-based capacity alerts and priority queues
  2. Define service-specific health SLOs and dependency diagrams
  3. Route incidents to the failing dependency based on correlated traces
  4. Treat every user timeout as identical model endpoint unavailability

Explanation

Correct Answer

B:

The sustained prevention measure is to define service-specific health SLOs and dependency diagrams. The implementation can still use the separate action to instrument each dependent Foundry and tool service with distinct success-rate metrics, but a guardrail constrains future changes. The manufacturing group can measure whether the rule works by arranging to simulate one service failure and verify the correct alert identifies it.

Incorrect Answers

A:

The action to establish token-based capacity alerts and priority queues addresses model request quota rather than the decisive provider health monitoring condition. The manufacturing group should not infer that fixing one adjacent AI service validates its unrelated workflow; the requested action is to define service-specific health SLOs and dependency diagrams.

C:

The action to route incidents to the failing dependency based on correlated traces can contribute to provider health monitoring as a corrective remediation step; here, an app needs to distinguish model failures from search and speech outages. For prevention at the manufacturing group, the measure to route incidents to the failing dependency based on correlated traces is weaker or aimed at another issue than the provider health monitoring guardrail to define service-specific health SLOs and dependency diagrams needed when an app needs to distinguish model failures from search and speech outages.

D:

For prevention at the manufacturing group, the measure to treat every user timeout as identical model endpoint unavailability is weaker or aimed at another issue than the provider health monitoring guardrail to define service-specific health SLOs and dependency diagrams needed when an app needs to distinguish model failures from search and speech outages. The lasting guardrail for this specific threat is to define service-specific health SLOs and dependency diagrams, not simply to complete an adjacent service task.

Question 200

During a service audit, a data engineer must accept a change for provider health monitoring. The issue to be addressed is: An app needs to distinguish model failures from search and speech outages. Which concrete verification action demonstrates that the relevant behavior works?

  1. Treat every user timeout as identical model endpoint unavailability
  2. Simulate one service failure and verify the correct alert identifies it
  3. Instrument each dependent Foundry and tool service with distinct success-rate metrics
  4. Evaluate a fixed representative prompt set across both versions

Explanation

Correct Answer

B:

To verify this change, the legal operations unit should simulate one service failure and verify the correct alert identifies it. Operators still need the implementation to instrument each dependent Foundry and tool service with distinct success-rate metrics and the preventive requirement to define service-specific health SLOs and dependency diagrams.

Incorrect Answers

A:

In the legal operations unit acceptance review, trying to treat every user timeout as identical model endpoint unavailability does not test the exact provider health monitoring condition where an app needs to distinguish model failures from search and speech outages; the relevant test is to simulate one service failure and verify the correct alert identifies it. Acceptance should instead demonstrate that the service can simulate one service failure and verify the correct alert identifies it, using the described failure as the test criterion.

C:

The action to instrument each dependent Foundry and tool service with distinct success-rate metrics can contribute to provider health monitoring as a implementation step; here, an app needs to distinguish model failures from search and speech outages. In the legal operations unit acceptance review, trying to instrument each dependent Foundry and tool service with distinct success-rate metrics does not test the exact provider health monitoring condition where an app needs to distinguish model failures from search and speech outages; the relevant test is to simulate one service failure and verify the correct alert identifies it.

D:

For model output drift, it can be sensible to evaluate a fixed representative prompt set across both versions if an assistant becomes less accurate after a deployment or prompt change. Without the actual change to simulate one service failure and verify the correct alert identifies it, the answer lacks the correct operational scope even if the alternate action is technically valid elsewhere.

Developing AI Apps and Agents on Azure (AI-103) Practice Tests