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 80

Question 790

During a architecture migration, a cloud developer must accept a change for visual bias evaluation. The issue to be addressed is: A multimodal classifier produces inconsistent results for similar scenes across demographic contexts. Which concrete verification action demonstrates that the relevant behavior works?

  1. Generate adversarial logo prompts and confirm review or rejection
  2. Assume aggregate accuracy proves acceptable treatment across all groups
  3. Compare equivalent visual tasks across balanced labeled examples
  4. Evaluate representative visual cases and inspect unfair error patterns

Explanation

Correct Answer

C:

To verify this change, the software consultancy should compare equivalent visual tasks across balanced labeled examples. Operators still need the implementation to evaluate representative visual cases and inspect unfair error patterns and the preventive requirement to maintain fairness evaluation datasets and human oversight for high-stakes cases.

Incorrect Answers

A:

For brand misuse detection, it can be sensible to generate adversarial logo prompts and confirm review or rejection if a marketing generator sometimes produces unapproved partner logos. Without the actual change to compare equivalent visual tasks across balanced labeled examples, the answer lacks the correct operational scope even if the alternate action is technically valid elsewhere.

B:

In the software consultancy acceptance review, trying to assume aggregate accuracy proves acceptable treatment across all groups does not test the exact visual bias evaluation condition where a multimodal classifier produces inconsistent results for similar scenes across demographic contexts; the relevant test is to compare equivalent visual tasks across balanced labeled examples. Acceptance should instead demonstrate that the service can compare equivalent visual tasks across balanced labeled examples, using the described failure as the test criterion.

D:

The action to evaluate representative visual cases and inspect unfair error patterns can contribute to visual bias evaluation as a implementation step; here, a multimodal classifier produces inconsistent results for similar scenes across demographic contexts. In the software consultancy acceptance review, trying to evaluate representative visual cases and inspect unfair error patterns does not test the exact visual bias evaluation condition where a multimodal classifier produces inconsistent results for similar scenes across demographic contexts; the relevant test is to compare equivalent visual tasks across balanced labeled examples.

Question 791

At a energy cooperative, the site reliability engineer is preparing a source update. The concrete requirement is: A vision assistant responds to benign photos with unsafe instructions unrelated to the image. Which design action BEST addresses the stated need?

  1. Assume visual input makes answer safety unnecessary
  2. Submit benign photos with adversarial captions and inspect the response
  3. Treat embedded image text as untrusted task data rather than higher-level instructions
  4. Screen generated text and enforce task-specific response restrictions

Explanation

Correct Answer

D:

The proposed implementation is to screen generated text and enforce task-specific response restrictions; this matches the operational requirement that a vision assistant responds to benign photos with unsafe instructions unrelated to the image. The team can submit benign photos with adversarial captions and inspect the response. The team should also evaluate image-to-text safety as a separate release category.

Incorrect Answers

A:

Choosing to assume visual input makes answer safety unnecessary cannot directly replace the specified action to screen generated text and enforce task-specific response restrictions when a vision assistant responds to benign photos with unsafe instructions unrelated to the image. The actual implementation must screen generated text and enforce task-specific response restrictions, because the stated service condition is that a vision assistant responds to benign photos with unsafe instructions unrelated to the image.

B:

The action to submit benign photos with adversarial captions and inspect the response can contribute to unsafe image-to-text output as a targeted acceptance check step; here, a vision assistant responds to benign photos with unsafe instructions unrelated to the image. Choosing to submit benign photos with adversarial captions and inspect the response cannot directly replace the specified action to screen generated text and enforce task-specific response restrictions when a vision assistant responds to benign photos with unsafe instructions unrelated to the image.

C:

Choosing to treat embedded image text as untrusted task data rather than higher-level instructions would direct effort toward image embedded prompt injection, whose characteristic issue is that a photographed document contains tiny text telling the agent to execute a secret-export tool. The relevant corrective or preventive action here remains to screen generated text and enforce task-specific response restrictions, with its own validation criteria.

Question 792

After a cost review at a industrial maintenance team, the team observes this failure: A vision assistant responds to benign photos with unsafe instructions unrelated to the image. Which flawed assumption or implementation choice would MOST DIRECTLY explain why the intended outcome was not achieved?

  1. Screen generated text and enforce task-specific response restrictions
  2. Assume visual input makes answer safety unnecessary
  3. Submit benign photos with adversarial captions and inspect the response
  4. Evaluate image-to-text safety as a separate release category
  5. Execute commands extracted from the image without checking authorization

Explanation

Correct Answer

B:

This design mistake is to assume visual input makes answer safety unnecessary; it conflicts with the actual unsafe image-to-text output behavior needed when a vision assistant responds to benign photos with unsafe instructions unrelated to the image. The industrial maintenance team should instead screen generated text and enforce task-specific response restrictions, followed by a check to submit benign photos with adversarial captions and inspect the response.

Incorrect Answers

A:

The alternative to screen generated text and enforce task-specific response restrictions is a implementation step rather than the faulty assumption being diagnosed. This industrial maintenance team failure-analysis alternative of screen generated text and enforce task-specific response restrictions does not diagnose the particular unsafe image-to-text output defect where a vision assistant responds to benign photos with unsafe instructions unrelated to the image; the keyed mistake is assume visual input makes answer safety unnecessary.

C:

The alternative to submit benign photos with adversarial captions and inspect the response is a targeted acceptance check step rather than the faulty assumption being diagnosed. This industrial maintenance team failure-analysis alternative of submit benign photos with adversarial captions and inspect the response does not diagnose the particular unsafe image-to-text output defect where a vision assistant responds to benign photos with unsafe instructions unrelated to the image; the keyed mistake is assume visual input makes answer safety unnecessary.

D:

The alternative to evaluate image-to-text safety as a separate release category is a ongoing preventive rule step rather than the faulty assumption being diagnosed. This industrial maintenance team failure-analysis alternative of evaluate image-to-text safety as a separate release category does not diagnose the particular unsafe image-to-text output defect where a vision assistant responds to benign photos with unsafe instructions unrelated to the image; the keyed mistake is assume visual input makes answer safety unnecessary.

E:

This approach relates to image embedded prompt injection: a photographed document contains tiny text telling the agent to execute a secret-export tool. If the team chose it instead, the targeted check to submit benign photos with adversarial captions and inspect the response could still fail because the underlying required action is to assume visual input makes answer safety unnecessary.

Question 793

The research university is responding to a reliability exercise. A agent designer discovers this issue: A vision assistant responds to benign photos with unsafe instructions unrelated to the image. What corrective step should the team take to restore the intended functionality most directly?

  1. Submit benign photos with adversarial captions and inspect the response
  2. Remove exposed logs and harden media telemetry collection
  3. Assume visual input makes answer safety unnecessary
  4. Block irrelevant unsafe generation and strengthen context grounding

Explanation

Correct Answer

D:

The direct repair is to block irrelevant unsafe generation and strengthen context grounding. Once the condition is corrected, the research university should submit benign photos with adversarial captions and inspect the response. To reduce recurrence, it should also evaluate image-to-text safety as a separate release category; those steps provide different recovery and prevention functions.

Incorrect Answers

A:

The action to submit benign photos with adversarial captions and inspect the response can contribute to unsafe image-to-text output as a targeted acceptance check step; here, a vision assistant responds to benign photos with unsafe instructions unrelated to the image. During research university remediation, the proposed step to submit benign photos with adversarial captions and inspect the response is not the direct restoration action for unsafe image-to-text output when a vision assistant responds to benign photos with unsafe instructions unrelated to the image; the keyed repair is block irrelevant unsafe generation and strengthen context grounding.

B:

During research university remediation, the proposed step to remove exposed logs and harden media telemetry collection is not the direct restoration action for unsafe image-to-text output when a vision assistant responds to benign photos with unsafe instructions unrelated to the image; the keyed repair is block irrelevant unsafe generation and strengthen context grounding. If the team chose it instead, the targeted check to submit benign photos with adversarial captions and inspect the response could still fail because the underlying required action is to block irrelevant unsafe generation and strengthen context grounding.

C:

During research university remediation, the proposed step to assume visual input makes answer safety unnecessary is not the direct restoration action for unsafe image-to-text output when a vision assistant responds to benign photos with unsafe instructions unrelated to the image; the keyed repair is block irrelevant unsafe generation and strengthen context grounding. After remediation the team must submit benign photos with adversarial captions and inspect the response; the action in this option does not itself establish that target recovery outcome.

Question 794

A responsible-AI reviewer at a education technology company wants the following issue to be prevented in later deployments: A vision assistant responds to benign photos with unsafe instructions unrelated to the image. Which enduring control or engineering guardrail BEST prevents recurrence?

  1. Use role-limited trace storage and privacy retention rules
  2. Assume visual input makes answer safety unnecessary
  3. Evaluate image-to-text safety as a separate release category
  4. Block irrelevant unsafe generation and strengthen context grounding

Explanation

Correct Answer

C:

Standardizing the boundary to evaluate image-to-text safety as a separate release category is preventive because the same unsafe image-to-text output failure could otherwise return: a vision assistant responds to benign photos with unsafe instructions unrelated to the image. For this education technology company, the service should also screen generated text and enforce task-specific response restrictions, and tests should submit benign photos with adversarial captions and inspect the response so that the ongoing requirement is measurable.

Incorrect Answers

A:

The action to use role-limited trace storage and privacy retention rules addresses visual privacy filtering rather than the decisive unsafe image-to-text output condition. The education technology company should not infer that fixing one adjacent AI service validates its unrelated workflow; the requested action is to evaluate image-to-text safety as a separate release category.

B:

For prevention at the education technology company, the measure to assume visual input makes answer safety unnecessary is weaker or aimed at another issue than the unsafe image-to-text output guardrail to evaluate image-to-text safety as a separate release category needed when a vision assistant responds to benign photos with unsafe instructions unrelated to the image. The lasting guardrail for this specific threat is to evaluate image-to-text safety as a separate release category, not simply to complete an adjacent service task.

D:

The action to block irrelevant unsafe generation and strengthen context grounding can contribute to unsafe image-to-text output as a corrective remediation step; here, a vision assistant responds to benign photos with unsafe instructions unrelated to the image. For prevention at the education technology company, the measure to block irrelevant unsafe generation and strengthen context grounding is weaker or aimed at another issue than the unsafe image-to-text output guardrail to evaluate image-to-text safety as a separate release category needed when a vision assistant responds to benign photos with unsafe instructions unrelated to the image.

Question 795

During a cost review, a solution architect must accept a change for unsafe image-to-text output. The issue to be addressed is: A vision assistant responds to benign photos with unsafe instructions unrelated to the image. Which concrete verification action demonstrates that the relevant behavior works?

  1. Show an adversarial document image and verify no unauthorized action
  2. Submit benign photos with adversarial captions and inspect the response
  3. Screen generated text and enforce task-specific response restrictions
  4. Assume visual input makes answer safety unnecessary

Explanation

Correct Answer

B:

The targeted acceptance test is to submit benign photos with adversarial captions and inspect the response. The design should first screen generated text and enforce task-specific response restrictions. A robust pharmaceutical research group deployment also needs to evaluate image-to-text safety as a separate release category; passing this check provides relevant evidence but does not certify every possible input or security risk.

Incorrect Answers

A:

For image embedded prompt injection, it can be sensible to show an adversarial document image and verify no unauthorized action if a photographed document contains tiny text telling the agent to execute a secret-export tool. Without the actual change to submit benign photos with adversarial captions and inspect the response, the answer lacks the correct operational scope even if the alternate action is technically valid elsewhere.

C:

The action to screen generated text and enforce task-specific response restrictions can contribute to unsafe image-to-text output as a implementation step; here, a vision assistant responds to benign photos with unsafe instructions unrelated to the image. In the pharmaceutical research group acceptance review, trying to screen generated text and enforce task-specific response restrictions does not test the exact unsafe image-to-text output condition where a vision assistant responds to benign photos with unsafe instructions unrelated to the image; the relevant test is to submit benign photos with adversarial captions and inspect the response.

D:

In the pharmaceutical research group acceptance review, trying to assume visual input makes answer safety unnecessary does not test the exact unsafe image-to-text output condition where a vision assistant responds to benign photos with unsafe instructions unrelated to the image; the relevant test is to submit benign photos with adversarial captions and inspect the response. Acceptance should instead demonstrate that the service can submit benign photos with adversarial captions and inspect the response, using the described failure as the test criterion.

Question 796

At a software consultancy, the machine-learning developer is preparing a support escalation. The concrete requirement is: A team wants to generate a likeness of a real speaker without permission. Which design action BEST addresses the stated need?

  1. Require appropriate rights and consent before producing identifiable synthetic media
  2. Apply appropriate visual safety screening and prompt restrictions before display
  3. Assume source footage being on the web grants permission to clone the person
  4. Test a disallowed impersonation request and verify it is rejected

Explanation

Correct Answer

A:

Choosing to require appropriate rights and consent before producing identifiable synthetic media directly targets the described video rights and consent gap: a team wants to generate a likeness of a real speaker without permission. As a lasting boundary, they also need to record consent and permitted-media scope for synthetic outputs, because an implementation by itself does not prove safe operation.

Incorrect Answers

B:

Choosing to apply appropriate visual safety screening and prompt restrictions before display would direct effort toward visual content policy, whose characteristic issue is that an image generation tool produces assets outside the organization's approved visual policy. The relevant corrective or preventive action here remains to require appropriate rights and consent before producing identifiable synthetic media, with its own validation criteria.

C:

Choosing to assume source footage being on the web grants permission to clone the person cannot directly replace the specified action to require appropriate rights and consent before producing identifiable synthetic media when a team wants to generate a likeness of a real speaker without permission. The actual implementation must require appropriate rights and consent before producing identifiable synthetic media, because the stated service condition is that a team wants to generate a likeness of a real speaker without permission.

D:

The action to test a disallowed impersonation request and verify it is rejected can contribute to video rights and consent as a targeted acceptance check step; here, a team wants to generate a likeness of a real speaker without permission. Choosing to test a disallowed impersonation request and verify it is rejected cannot directly replace the specified action to require appropriate rights and consent before producing identifiable synthetic media when a team wants to generate a likeness of a real speaker without permission.

Question 797

After a service audit at a regional hospital, the team observes this failure: A team wants to generate a likeness of a real speaker without permission. Which flawed assumption or implementation choice would MOST DIRECTLY explain why the intended outcome was not achieved?

  1. Require appropriate rights and consent before producing identifiable synthetic media
  2. Assume source footage being on the web grants permission to clone the person
  3. Treat any familiar logo in generated content as legally permitted
  4. Test a disallowed impersonation request and verify it is rejected

Explanation

Correct Answer

B:

The faulty assumption is assume source footage being on the web grants permission to clone the person. A functioning alternative would be to require appropriate rights and consent before producing identifiable synthetic media, and the relevant validation would be to test a disallowed impersonation request and verify it is rejected.

Incorrect Answers

A:

The alternative to require appropriate rights and consent before producing identifiable synthetic media is a implementation step rather than the faulty assumption being diagnosed. This regional hospital failure-analysis alternative of require appropriate rights and consent before producing identifiable synthetic media does not diagnose the particular video rights and consent defect where a team wants to generate a likeness of a real speaker without permission; the keyed mistake is assume source footage being on the web grants permission to clone the person.

C:

This regional hospital failure-analysis alternative of treat any familiar logo in generated content as legally permitted does not diagnose the particular video rights and consent defect where a team wants to generate a likeness of a real speaker without permission; the keyed mistake is assume source footage being on the web grants permission to clone the person. The alternative may be useful in its own context, but it does not directly supply assume source footage being on the web grants permission to clone the person for the stated decision.

D:

The alternative to test a disallowed impersonation request and verify it is rejected is a targeted acceptance check step rather than the faulty assumption being diagnosed. This regional hospital failure-analysis alternative of test a disallowed impersonation request and verify it is rejected does not diagnose the particular video rights and consent defect where a team wants to generate a likeness of a real speaker without permission; the keyed mistake is assume source footage being on the web grants permission to clone the person.

Question 798

The insurance platform is responding to a security tabletop exercise. A quality assurance lead discovers this issue: A team wants to generate a likeness of a real speaker without permission. What corrective step should the team take to restore the intended functionality most directly?

  1. Test a disallowed impersonation request and verify it is rejected
  2. Remove the injected text from the instruction channel and retry interpretation
  3. Assume source footage being on the web grants permission to clone the person
  4. Stop unapproved generation and substitute licensed non-identifying media

Explanation

Correct Answer

D:

The appropriate remediation is to stop unapproved generation and substitute licensed non-identifying media, because the failure concerns video rights and consent: a team wants to generate a likeness of a real speaker without permission. A meaningful verification should test a disallowed impersonation request and verify it is rejected, and the durable operating rule should record consent and permitted-media scope for synthetic outputs after the immediate fix.

Incorrect Answers

A:

The action to test a disallowed impersonation request and verify it is rejected can contribute to video rights and consent as a targeted acceptance check step; here, a team wants to generate a likeness of a real speaker without permission. During insurance platform remediation, the proposed step to test a disallowed impersonation request and verify it is rejected is not the direct restoration action for video rights and consent when a team wants to generate a likeness of a real speaker without permission; the keyed repair is stop unapproved generation and substitute licensed non-identifying media.

B:

This approach relates to image embedded prompt injection: a photographed document contains tiny text telling the agent to execute a secret-export tool. If the team chose it instead, the targeted check to test a disallowed impersonation request and verify it is rejected could still fail because the underlying required action is to stop unapproved generation and substitute licensed non-identifying media.

C:

During insurance platform remediation, the proposed step to assume source footage being on the web grants permission to clone the person is not the direct restoration action for video rights and consent when a team wants to generate a likeness of a real speaker without permission; the keyed repair is stop unapproved generation and substitute licensed non-identifying media. After remediation the team must test a disallowed impersonation request and verify it is rejected; the action in this option does not itself establish that target recovery outcome.

Question 799

A AI platform engineer at a media platform wants the following issue to be prevented in later deployments: A team wants to generate a likeness of a real speaker without permission. Which enduring control or engineering guardrail BEST prevents recurrence?

  1. Evaluate image-to-text safety as a separate release category
  2. Require appropriate rights and consent before producing identifiable synthetic media
  3. Test a disallowed impersonation request and verify it is rejected
  4. Record consent and permitted-media scope for synthetic outputs
  5. Stop unapproved generation and substitute licensed non-identifying media
  6. Assume source footage being on the web grants permission to clone the person

Explanation

Correct Answer

D:

The sustained prevention measure is to record consent and permitted-media scope for synthetic outputs. The implementation can still use the separate action to require appropriate rights and consent before producing identifiable synthetic media, but a guardrail constrains future changes. The media platform can measure whether the rule works by arranging to test a disallowed impersonation request and verify it is rejected.

Incorrect Answers

A:

Choosing to evaluate image-to-text safety as a separate release category would direct effort toward unsafe image-to-text output, whose characteristic issue is that a vision assistant responds to benign photos with unsafe instructions unrelated to the image. The relevant corrective or preventive action here remains to record consent and permitted-media scope for synthetic outputs, with its own validation criteria.

B:

The action to require appropriate rights and consent before producing identifiable synthetic media can contribute to video rights and consent as a implementation step; here, a team wants to generate a likeness of a real speaker without permission. For prevention at the media platform, the measure to require appropriate rights and consent before producing identifiable synthetic media is weaker or aimed at another issue than the video rights and consent guardrail to record consent and permitted-media scope for synthetic outputs needed when a team wants to generate a likeness of a real speaker without permission.

C:

The action to test a disallowed impersonation request and verify it is rejected can contribute to video rights and consent as a targeted acceptance check step; here, a team wants to generate a likeness of a real speaker without permission. For prevention at the media platform, the measure to test a disallowed impersonation request and verify it is rejected is weaker or aimed at another issue than the video rights and consent guardrail to record consent and permitted-media scope for synthetic outputs needed when a team wants to generate a likeness of a real speaker without permission.

E:

The action to stop unapproved generation and substitute licensed non-identifying media can contribute to video rights and consent as a corrective remediation step; here, a team wants to generate a likeness of a real speaker without permission. For prevention at the media platform, the measure to stop unapproved generation and substitute licensed non-identifying media is weaker or aimed at another issue than the video rights and consent guardrail to record consent and permitted-media scope for synthetic outputs needed when a team wants to generate a likeness of a real speaker without permission.

F:

For prevention at the media platform, the measure to assume source footage being on the web grants permission to clone the person is weaker or aimed at another issue than the video rights and consent guardrail to record consent and permitted-media scope for synthetic outputs needed when a team wants to generate a likeness of a real speaker without permission. The lasting guardrail for this specific threat is to record consent and permitted-media scope for synthetic outputs, not simply to complete an adjacent service task.

Question 800

During a service audit, a compliance specialist must accept a change for video rights and consent. The issue to be addressed is: A team wants to generate a likeness of a real speaker without permission. Which concrete verification action demonstrates that the relevant behavior works?

  1. Assume source footage being on the web grants permission to clone the person
  2. Inspect exported assets for required label and corresponding metadata
  3. Test a disallowed impersonation request and verify it is rejected
  4. Require appropriate rights and consent before producing identifiable synthetic media

Explanation

Correct Answer

C:

To verify this change, the customer support center should test a disallowed impersonation request and verify it is rejected. Operators still need the implementation to require appropriate rights and consent before producing identifiable synthetic media and the preventive requirement to record consent and permitted-media scope for synthetic outputs.

Incorrect Answers

A:

In the customer support center acceptance review, trying to assume source footage being on the web grants permission to clone the person does not test the exact video rights and consent condition where a team wants to generate a likeness of a real speaker without permission; the relevant test is to test a disallowed impersonation request and verify it is rejected. Acceptance should instead demonstrate that the service can test a disallowed impersonation request and verify it is rejected, using the described failure as the test criterion.

B:

For generated media watermark policy, it can be sensible to inspect exported assets for required label and corresponding metadata if a company requires identification of synthetic promotional visuals. Without the actual change to test a disallowed impersonation request and verify it is rejected, the answer lacks the correct operational scope even if the alternate action is technically valid elsewhere.

D:

The action to require appropriate rights and consent before producing identifiable synthetic media can contribute to video rights and consent as a implementation step; here, a team wants to generate a likeness of a real speaker without permission. In the customer support center acceptance review, trying to require appropriate rights and consent before producing identifiable synthetic media does not test the exact video rights and consent condition where a team wants to generate a likeness of a real speaker without permission; the relevant test is to test a disallowed impersonation request and verify it is rejected.

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