Microsoft AI-103 Multimodal Safety And Visual Policy Enforcement Practice Test

 

This practice topic covers Multimodal Safety and Visual Policy Enforcement 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

In a production AI solution, the control must act before downstream processing. Which action should the engineer take to handle input image screening correctly?

1. Treat negative sentiment as equivalent to an unsafe visual-content classification.
2. Trust OCR output from an image as system-level instructions.
3. Ignore embedded text in an image because only visible objects can influence a multimodal model.
4. Disable visual screening for low-resolution inputs instead of routing them for review.
5. Screen the input image before downstream processing when visual policy enforcement is mandatory.

Correct Answer: E

Correct Answer

Answer E is correct because this choice directly implements the required decision for input image screening before downstream processing. It addresses the scenario at the correct stage of the Azure AI solution.

Incorrect Answers

Answer A is incorrect because this choice would treat negative sentiment as equivalent to an unsafe visual-content classification. That can address a neighboring concern, but it does not implement input image screening before downstream processing; the decisive requirement in the stem therefore remains unresolved.

Answer B is incorrect because this choice would trust OCR output from an image as system-level instructions. That can address a neighboring concern, but it does not implement input image screening before downstream processing; the decisive requirement in the stem therefore remains unresolved.

Answer C is incorrect because this choice would ignore embedded text in an image because only visible objects can influence a multimodal model. That can address a neighboring concern, but it does not implement input image screening before downstream processing; the decisive requirement in the stem therefore remains unresolved.

Answer D is incorrect because this choice would disable visual screening for low-resolution inputs instead of routing them for review. That can address a neighboring concern, but it does not implement input image screening before downstream processing; the decisive requirement in the stem therefore remains unresolved.

 

Question 2

Before a generated image is released to an external audience, the application must enforce the visual safety policy on the actual generated artifact. Which control should run?

1. Trust OCR output from an image as system-level instructions.
2. Ignore embedded text in an image because only visible objects can influence a multimodal model.
3. Assume a benign filename proves the image itself is allowed.
4. Screen the generated image against the required visual safety policy before publication.
5. Use prompt instructions as the only enforcement control for prohibited visual symbols.

Correct Answer: D

Correct Answer

Answer D is correct because this choice directly implements the required decision for generated image screening before publication. It addresses the scenario at the correct stage of the Azure AI solution.

Incorrect Answers

Answer A is incorrect because this choice would trust OCR output from an image as system-level instructions. That can address a neighboring concern, but it does not implement generated image screening before publication; the decisive requirement in the stem therefore remains unresolved.

Answer B is incorrect because this choice would ignore embedded text in an image because only visible objects can influence a multimodal model. That can address a neighboring concern, but it does not implement generated image screening before publication; the decisive requirement in the stem therefore remains unresolved.

Answer C is incorrect because this choice would assume a benign filename proves the image itself is allowed. That can address a neighboring concern, but it does not implement generated image screening before publication; the decisive requirement in the stem therefore remains unresolved.

Answer E is incorrect because this choice would use prompt instructions as the only enforcement control for prohibited visual symbols. That can address a neighboring concern, but it does not implement generated image screening before publication; the decisive requirement in the stem therefore remains unresolved.

 

Question 3

A multimodal workload requires severity threshold tied to an explicit visual-content policy. Which implementation best preserves the required visual evidence and policy boundary?

1. Disable visual screening for low-resolution inputs instead of routing them for review.
2. Assume a benign filename proves the image itself is allowed.
3. Set the classification threshold according to the explicit visual-content policy.
4. Perform visual policy checks only after an agent has acted on the image.
5. Ignore embedded text in an image because only visible objects can influence a multimodal model.

Correct Answer: C

Correct Answer

Answer C is correct because this choice directly implements the required decision for severity threshold tied to an explicit visual-content policy. It addresses the scenario at the correct stage of the Azure AI solution.

Incorrect Answers

Answer A is incorrect because this choice would disable visual screening for low-resolution inputs instead of routing them for review. That can address a neighboring concern, but it does not implement severity threshold tied to an explicit visual-content policy; the decisive requirement in the stem therefore remains unresolved.

Answer B is incorrect because this choice would assume a benign filename proves the image itself is allowed. That can address a neighboring concern, but it does not implement severity threshold tied to an explicit visual-content policy; the decisive requirement in the stem therefore remains unresolved.

Answer D is incorrect because this choice would perform visual policy checks only after an agent has acted on the image. That can address a neighboring concern, but it does not implement severity threshold tied to an explicit visual-content policy; the decisive requirement in the stem therefore remains unresolved.

Answer E is incorrect because this choice would ignore embedded text in an image because only visible objects can influence a multimodal model. That can address a neighboring concern, but it does not implement severity threshold tied to an explicit visual-content policy; the decisive requirement in the stem therefore remains unresolved.

 

Question 4

In a production AI solution, the required outcome is human review. Which action should the engineer take to handle borderline classification routed correctly?

1. Ignore embedded text in an image because only visible objects can influence a multimodal model.
2. Route borderline classifications for human review.
3. Perform visual policy checks only after an agent has acted on the image.
4. Treat negative sentiment as equivalent to an unsafe visual-content classification.
5. Use prompt instructions as the only enforcement control for prohibited visual symbols.

Correct Answer: B

Correct Answer

Answer B is correct because this choice directly implements the required decision for borderline classification routed for human review. It addresses the scenario at the correct stage of the Azure AI solution.

Incorrect Answers

Answer A is incorrect because this choice would ignore embedded text in an image because only visible objects can influence a multimodal model. That can address a neighboring concern, but it does not implement borderline classification routed for human review; the decisive requirement in the stem therefore remains unresolved.

Answer C is incorrect because this choice would perform visual policy checks only after an agent has acted on the image. That can address a neighboring concern, but it does not implement borderline classification routed for human review; the decisive requirement in the stem therefore remains unresolved.

Answer D is incorrect because this choice would treat negative sentiment as equivalent to an unsafe visual-content classification. That can address a neighboring concern, but it does not implement borderline classification routed for human review; the decisive requirement in the stem therefore remains unresolved.

Answer E is incorrect because this choice would use prompt instructions as the only enforcement control for prohibited visual symbols. That can address a neighboring concern, but it does not implement borderline classification routed for human review; the decisive requirement in the stem therefore remains unresolved.

 

Question 5

A multimodal workload requires separate handling of disallowed content and uncertain classification. Which implementation best preserves the required visual evidence and policy boundary?

1. Handle clearly disallowed content differently from uncertain classifications that require review.
2. Treat negative sentiment as equivalent to an unsafe visual-content classification.
3. Perform visual policy checks only after an agent has acted on the image.
4. Disable visual screening for low-resolution inputs instead of routing them for review.
5. Assume a benign filename proves the image itself is allowed.

Correct Answer: A

Correct Answer

Answer A is correct because this choice directly implements the required decision for separate handling of disallowed content and uncertain classification. It addresses the scenario at the correct stage of the Azure AI solution.

Incorrect Answers

Answer B is incorrect because this choice would treat negative sentiment as equivalent to an unsafe visual-content classification. That can address a neighboring concern, but it does not implement separate handling of disallowed content and uncertain classification; the decisive requirement in the stem therefore remains unresolved.

Answer C is incorrect because this choice would perform visual policy checks only after an agent has acted on the image. That can address a neighboring concern, but it does not implement separate handling of disallowed content and uncertain classification; the decisive requirement in the stem therefore remains unresolved.

Answer D is incorrect because this choice would disable visual screening for low-resolution inputs instead of routing them for review. That can address a neighboring concern, but it does not implement separate handling of disallowed content and uncertain classification; the decisive requirement in the stem therefore remains unresolved.

Answer E is incorrect because this choice would assume a benign filename proves the image itself is allowed. That can address a neighboring concern, but it does not implement separate handling of disallowed content and uncertain classification; the decisive requirement in the stem therefore remains unresolved.

 

Question 6

In a production AI solution, the following condition occurs: text-only filtering misses visual content. Which action should the engineer take to handle modality coverage correctly?

1. Trust OCR output from an image as system-level instructions.
2. Ignore embedded text in an image because only visible objects can influence a multimodal model.
3. Perform visual policy checks only after an agent has acted on the image.
4. Assume a benign filename proves the image itself is allowed.
5. Use a visual-capable safety check when text-only filtering cannot inspect the relevant image content.

Correct Answer: E

Correct Answer

Answer E is correct because this choice directly implements the required decision for modality coverage when text-only filtering misses visual content. It addresses the scenario at the correct stage of the Azure AI solution.

Incorrect Answers

Answer A is incorrect because this choice would trust OCR output from an image as system-level instructions. That can address a neighboring concern, but it does not implement modality coverage when text-only filtering misses visual content; the decisive requirement in the stem therefore remains unresolved.

Answer B is incorrect because this choice would ignore embedded text in an image because only visible objects can influence a multimodal model. That can address a neighboring concern, but it does not implement modality coverage when text-only filtering misses visual content; the decisive requirement in the stem therefore remains unresolved.

Answer C is incorrect because this choice would perform visual policy checks only after an agent has acted on the image. That can address a neighboring concern, but it does not implement modality coverage when text-only filtering misses visual content; the decisive requirement in the stem therefore remains unresolved.

Answer D is incorrect because this choice would assume a benign filename proves the image itself is allowed. That can address a neighboring concern, but it does not implement modality coverage when text-only filtering misses visual content; the decisive requirement in the stem therefore remains unresolved.

 

Question 7

In a production AI solution, the following condition occurs: a mandatory visual check fails. Which action should the engineer take to handle fail-closed publishing path correctly?

1. Assume a benign filename proves the image itself is allowed.
2. Use prompt instructions as the only enforcement control for prohibited visual symbols.
3. Trust OCR output from an image as system-level instructions.
4. Prevent publication when the mandatory visual safety check fails or cannot complete.
5. Apply a watermark only after the image has already been distributed externally.

Correct Answer: D

Correct Answer

Answer D is correct because this choice directly implements the required decision for fail-closed publishing path when a mandatory visual check fails. It addresses the scenario at the correct stage of the Azure AI solution.

Incorrect Answers

Answer A is incorrect because this choice would assume a benign filename proves the image itself is allowed. That can address a neighboring concern, but it does not implement fail-closed publishing path when a mandatory visual check fails; the decisive requirement in the stem therefore remains unresolved.

Answer B is incorrect because this choice would use prompt instructions as the only enforcement control for prohibited visual symbols. That can address a neighboring concern, but it does not implement fail-closed publishing path when a mandatory visual check fails; the decisive requirement in the stem therefore remains unresolved.

Answer C is incorrect because this choice would trust OCR output from an image as system-level instructions. That can address a neighboring concern, but it does not implement fail-closed publishing path when a mandatory visual check fails; the decisive requirement in the stem therefore remains unresolved.

Answer E is incorrect because this choice would apply a watermark only after the image has already been distributed externally. That can address a neighboring concern, but it does not implement fail-closed publishing path when a mandatory visual check fails; the decisive requirement in the stem therefore remains unresolved.

 

Question 8

A multimodal workload requires embedded text treated as untrusted image content. Which implementation best preserves the required visual evidence and policy boundary?

1. Treat negative sentiment as equivalent to an unsafe visual-content classification.
2. Disable visual screening for low-resolution inputs instead of routing them for review.
3. Treat text embedded in an image as untrusted content.
4. Assume a benign filename proves the image itself is allowed.
5. Apply a watermark only after the image has already been distributed externally.

Correct Answer: C

Correct Answer

Answer C is correct because this choice directly implements the required decision for embedded text treated as untrusted image content. It addresses the scenario at the correct stage of the Azure AI solution.

Incorrect Answers

Answer A is incorrect because this choice would treat negative sentiment as equivalent to an unsafe visual-content classification. That can address a neighboring concern, but it does not implement embedded text treated as untrusted image content; the decisive requirement in the stem therefore remains unresolved.

Answer B is incorrect because this choice would disable visual screening for low-resolution inputs instead of routing them for review. That can address a neighboring concern, but it does not implement embedded text treated as untrusted image content; the decisive requirement in the stem therefore remains unresolved.

Answer D is incorrect because this choice would assume a benign filename proves the image itself is allowed. That can address a neighboring concern, but it does not implement embedded text treated as untrusted image content; the decisive requirement in the stem therefore remains unresolved.

Answer E is incorrect because this choice would apply a watermark only after the image has already been distributed externally. That can address a neighboring concern, but it does not implement embedded text treated as untrusted image content; the decisive requirement in the stem therefore remains unresolved.

 

Question 9

A multimodal workload requires oCR output prevented from overriding system instructions. Which implementation best preserves the required visual evidence and policy boundary?

1. Disable visual screening for low-resolution inputs instead of routing them for review.
2. Treat OCR text from an image as untrusted content and prevent it from overriding system instructions.
3. Use prompt instructions as the only enforcement control for prohibited visual symbols.
4. Perform visual policy checks only after an agent has acted on the image.
5. Trust OCR output from an image as system-level instructions.

Correct Answer: B

Correct Answer

Answer B is correct because this choice directly implements the required decision for oCR output prevented from overriding system instructions. It addresses the scenario at the correct stage of the Azure AI solution.

Incorrect Answers

Answer A is incorrect because this choice would disable visual screening for low-resolution inputs instead of routing them for review. That can address a neighboring concern, but it does not implement oCR output prevented from overriding system instructions; the decisive requirement in the stem therefore remains unresolved.

Answer C is incorrect because this choice would use prompt instructions as the only enforcement control for prohibited visual symbols. That can address a neighboring concern, but it does not implement oCR output prevented from overriding system instructions; the decisive requirement in the stem therefore remains unresolved.

Answer D is incorrect because this choice would perform visual policy checks only after an agent has acted on the image. That can address a neighboring concern, but it does not implement oCR output prevented from overriding system instructions; the decisive requirement in the stem therefore remains unresolved.

Answer E is incorrect because this choice would trust OCR output from an image as system-level instructions. That can address a neighboring concern, but it does not implement oCR output prevented from overriding system instructions; the decisive requirement in the stem therefore remains unresolved.

 

Question 10

A multimodal workload requires image-borne request to invoke an unauthorized tool. Which implementation best preserves the required visual evidence and policy boundary?

1. Do not allow image-borne instructions to invoke an unauthorized tool.
2. Use prompt instructions as the only enforcement control for prohibited visual symbols.
3. Treat negative sentiment as equivalent to an unsafe visual-content classification.
4. Perform visual policy checks only after an agent has acted on the image.
5. Trust OCR output from an image as system-level instructions.

Correct Answer: A

Correct Answer

Answer A is correct because this choice directly implements the required decision for image-borne request to invoke an unauthorized tool. It addresses the scenario at the correct stage of the Azure AI solution.

Incorrect Answers

Answer B is incorrect because this choice would use prompt instructions as the only enforcement control for prohibited visual symbols. That can address a neighboring concern, but it does not implement image-borne request to invoke an unauthorized tool; the decisive requirement in the stem therefore remains unresolved.

Answer C is incorrect because this choice would treat negative sentiment as equivalent to an unsafe visual-content classification. That can address a neighboring concern, but it does not implement image-borne request to invoke an unauthorized tool; the decisive requirement in the stem therefore remains unresolved.

Answer D is incorrect because this choice would perform visual policy checks only after an agent has acted on the image. That can address a neighboring concern, but it does not implement image-borne request to invoke an unauthorized tool; the decisive requirement in the stem therefore remains unresolved.

Answer E is incorrect because this choice would trust OCR output from an image as system-level instructions. That can address a neighboring concern, but it does not implement image-borne request to invoke an unauthorized tool; the decisive requirement in the stem therefore remains unresolved.

 

Question 11

A multimodal workload requires visual injection carried through retrieved document figures. Which implementation best preserves the required visual evidence and policy boundary?

1. Use prompt instructions as the only enforcement control for prohibited visual symbols.
2. Apply a watermark only after the image has already been distributed externally.
3. Treat negative sentiment as equivalent to an unsafe visual-content classification.
4. Assume a benign filename proves the image itself is allowed.
5. Treat instructions carried through retrieved figures as untrusted data.

Correct Answer: E

Correct Answer

Answer E is correct because this choice directly implements the required decision for visual injection carried through retrieved document figures. It addresses the scenario at the correct stage of the Azure AI solution.

Incorrect Answers

Answer A is incorrect because this choice would use prompt instructions as the only enforcement control for prohibited visual symbols. That can address a neighboring concern, but it does not implement visual injection carried through retrieved document figures; the decisive requirement in the stem therefore remains unresolved.

Answer B is incorrect because this choice would apply a watermark only after the image has already been distributed externally. That can address a neighboring concern, but it does not implement visual injection carried through retrieved document figures; the decisive requirement in the stem therefore remains unresolved.

Answer C is incorrect because this choice would treat negative sentiment as equivalent to an unsafe visual-content classification. That can address a neighboring concern, but it does not implement visual injection carried through retrieved document figures; the decisive requirement in the stem therefore remains unresolved.

Answer D is incorrect because this choice would assume a benign filename proves the image itself is allowed. That can address a neighboring concern, but it does not implement visual injection carried through retrieved document figures; the decisive requirement in the stem therefore remains unresolved.

 

Question 12

A multimodal workload requires multimodal attack test that harmless image text does not cover. Which implementation best preserves the required visual evidence and policy boundary?

1. Treat negative sentiment as equivalent to an unsafe visual-content classification.
2. Perform visual policy checks only after an agent has acted on the image.
3. Disable visual screening for low-resolution inputs instead of routing them for review.
4. Test attacks that specifically use image content or embedded text, not only text prompts.
5. Ignore embedded text in an image because only visible objects can influence a multimodal model.

Correct Answer: D

Correct Answer

Answer D is correct because this choice directly implements the required decision for multimodal attack test that harmless image text does not cover. It addresses the scenario at the correct stage of the Azure AI solution.

Incorrect Answers

Answer A is incorrect because this choice would treat negative sentiment as equivalent to an unsafe visual-content classification. That can address a neighboring concern, but it does not implement multimodal attack test that harmless image text does not cover; the decisive requirement in the stem therefore remains unresolved.

Answer B is incorrect because this choice would perform visual policy checks only after an agent has acted on the image. That can address a neighboring concern, but it does not implement multimodal attack test that harmless image text does not cover; the decisive requirement in the stem therefore remains unresolved.

Answer C is incorrect because this choice would disable visual screening for low-resolution inputs instead of routing them for review. That can address a neighboring concern, but it does not implement multimodal attack test that harmless image text does not cover; the decisive requirement in the stem therefore remains unresolved.

Answer E is incorrect because this choice would ignore embedded text in an image because only visible objects can influence a multimodal model. That can address a neighboring concern, but it does not implement multimodal attack test that harmless image text does not cover; the decisive requirement in the stem therefore remains unresolved.

 

Question 13

In a production AI solution, the following condition occurs: attack detection alone is insufficient. Which action should the engineer take to handle layered mitigation correctly?

1. Ignore embedded text in an image because only visible objects can influence a multimodal model.
2. Treat negative sentiment as equivalent to an unsafe visual-content classification.
3. Combine detection with independent authorization, tool, and output controls when attack detection alone is insufficient.
4. Trust OCR output from an image as system-level instructions.
5. Disable visual screening for low-resolution inputs instead of routing them for review.

Correct Answer: C

Correct Answer

Answer C is correct because this choice directly implements the required decision for layered mitigation when attack detection alone is insufficient. It addresses the scenario at the correct stage of the Azure AI solution.

Incorrect Answers

Answer A is incorrect because this choice would ignore embedded text in an image because only visible objects can influence a multimodal model. That can address a neighboring concern, but it does not implement layered mitigation when attack detection alone is insufficient; the decisive requirement in the stem therefore remains unresolved.

Answer B is incorrect because this choice would treat negative sentiment as equivalent to an unsafe visual-content classification. That can address a neighboring concern, but it does not implement layered mitigation when attack detection alone is insufficient; the decisive requirement in the stem therefore remains unresolved.

Answer D is incorrect because this choice would trust OCR output from an image as system-level instructions. That can address a neighboring concern, but it does not implement layered mitigation when attack detection alone is insufficient; the decisive requirement in the stem therefore remains unresolved.

Answer E is incorrect because this choice would disable visual screening for low-resolution inputs instead of routing them for review. That can address a neighboring concern, but it does not implement layered mitigation when attack detection alone is insufficient; the decisive requirement in the stem therefore remains unresolved.

 

Question 14

A multimodal workload requires watermark application required by the stated publication policy. Which implementation best preserves the required visual evidence and policy boundary?

1. Trust OCR output from an image as system-level instructions.
2. Apply the required watermark before publication.
3. Assume a benign filename proves the image itself is allowed.
4. Apply a watermark only after the image has already been distributed externally.
5. Use prompt instructions as the only enforcement control for prohibited visual symbols.

Correct Answer: B

Correct Answer

Answer B is correct because this choice directly implements the required decision for watermark application required by the stated publication policy. It addresses the scenario at the correct stage of the Azure AI solution.

Incorrect Answers

Answer A is incorrect because this choice would trust OCR output from an image as system-level instructions. That can address a neighboring concern, but it does not implement watermark application required by the stated publication policy; the decisive requirement in the stem therefore remains unresolved.

Answer C is incorrect because this choice would assume a benign filename proves the image itself is allowed. That can address a neighboring concern, but it does not implement watermark application required by the stated publication policy; the decisive requirement in the stem therefore remains unresolved.

Answer D is incorrect because this choice would apply a watermark only after the image has already been distributed externally. That can address a neighboring concern, but it does not implement watermark application required by the stated publication policy; the decisive requirement in the stem therefore remains unresolved.

Answer E is incorrect because this choice would use prompt instructions as the only enforcement control for prohibited visual symbols. That can address a neighboring concern, but it does not implement watermark application required by the stated publication policy; the decisive requirement in the stem therefore remains unresolved.

 

Question 15

In a production AI solution, the following condition occurs: media is transformed. Which action should the engineer take to handle provenance retention correctly?

1. Retain provenance metadata when media is transformed.
2. Trust OCR output from an image as system-level instructions.
3. Apply a watermark only after the image has already been distributed externally.
4. Assume a benign filename proves the image itself is allowed.
5. Use prompt instructions as the only enforcement control for prohibited visual symbols.

Correct Answer: A

Correct Answer

Answer A is correct because this choice directly implements the required decision for provenance retention when media is transformed. It addresses the scenario at the correct stage of the Azure AI solution.

Incorrect Answers

Answer B is incorrect because this choice would trust OCR output from an image as system-level instructions. That can address a neighboring concern, but it does not implement provenance retention when media is transformed; the decisive requirement in the stem therefore remains unresolved.

Answer C is incorrect because this choice would apply a watermark only after the image has already been distributed externally. That can address a neighboring concern, but it does not implement provenance retention when media is transformed; the decisive requirement in the stem therefore remains unresolved.

Answer D is incorrect because this choice would assume a benign filename proves the image itself is allowed. That can address a neighboring concern, but it does not implement provenance retention when media is transformed; the decisive requirement in the stem therefore remains unresolved.

Answer E is incorrect because this choice would use prompt instructions as the only enforcement control for prohibited visual symbols. That can address a neighboring concern, but it does not implement provenance retention when media is transformed; the decisive requirement in the stem therefore remains unresolved.

 

Question 16

A multimodal workload requires prohibited-symbol rule beyond generic harm categories. Which implementation best preserves the required visual evidence and policy boundary?

1. Disable visual screening for low-resolution inputs instead of routing them for review.
2. Use prompt instructions as the only enforcement control for prohibited visual symbols.
3. Perform visual policy checks only after an agent has acted on the image.
4. Treat negative sentiment as equivalent to an unsafe visual-content classification.
5. Apply the explicit prohibited-symbol rule in addition to generic harm classifiers.

Correct Answer: E

Correct Answer

Answer E is correct because this choice directly implements the required decision for prohibited-symbol rule beyond generic harm categories. It addresses the scenario at the correct stage of the Azure AI solution.

Incorrect Answers

Answer A is incorrect because this choice would disable visual screening for low-resolution inputs instead of routing them for review. That can address a neighboring concern, but it does not implement prohibited-symbol rule beyond generic harm categories; the decisive requirement in the stem therefore remains unresolved.

Answer B is incorrect because this choice would use prompt instructions as the only enforcement control for prohibited visual symbols. That can address a neighboring concern, but it does not implement prohibited-symbol rule beyond generic harm categories; the decisive requirement in the stem therefore remains unresolved.

Answer C is incorrect because this choice would perform visual policy checks only after an agent has acted on the image. That can address a neighboring concern, but it does not implement prohibited-symbol rule beyond generic harm categories; the decisive requirement in the stem therefore remains unresolved.

Answer D is incorrect because this choice would treat negative sentiment as equivalent to an unsafe visual-content classification. That can address a neighboring concern, but it does not implement prohibited-symbol rule beyond generic harm categories; the decisive requirement in the stem therefore remains unresolved.

 

Question 17

In a production AI solution, the workload operates under an explicit permission rule. Which action should the engineer take to handle brand asset usage correctly?

1. Treat negative sentiment as equivalent to an unsafe visual-content classification.
2. Apply a watermark only after the image has already been distributed externally.
3. Ignore embedded text in an image because only visible objects can influence a multimodal model.
4. Enforce the stated permission rule before using protected brand assets under an explicit permission rule.
5. Trust OCR output from an image as system-level instructions.

Correct Answer: D

Correct Answer

Answer D is correct because this choice directly implements the required decision for brand asset usage under an explicit permission rule. It addresses the scenario at the correct stage of the Azure AI solution.

Incorrect Answers

Answer A is incorrect because this choice would treat negative sentiment as equivalent to an unsafe visual-content classification. That can address a neighboring concern, but it does not implement brand asset usage under an explicit permission rule; the decisive requirement in the stem therefore remains unresolved.

Answer B is incorrect because this choice would apply a watermark only after the image has already been distributed externally. That can address a neighboring concern, but it does not implement brand asset usage under an explicit permission rule; the decisive requirement in the stem therefore remains unresolved.

Answer C is incorrect because this choice would ignore embedded text in an image because only visible objects can influence a multimodal model. That can address a neighboring concern, but it does not implement brand asset usage under an explicit permission rule; the decisive requirement in the stem therefore remains unresolved.

Answer E is incorrect because this choice would trust OCR output from an image as system-level instructions. That can address a neighboring concern, but it does not implement brand asset usage under an explicit permission rule; the decisive requirement in the stem therefore remains unresolved.

 

Question 18

In a production AI solution, the required outcome is potentially inappropriate imagery. Which action should the engineer take to handle review workflow correctly?

1. Assume a benign filename proves the image itself is allowed.
2. Perform visual policy checks only after an agent has acted on the image.
3. Route potentially inappropriate imagery through the required review workflow.
4. Trust OCR output from an image as system-level instructions.
5. Apply a watermark only after the image has already been distributed externally.

Correct Answer: C

Correct Answer

Answer C is correct because this choice directly implements the required decision for review workflow for potentially inappropriate imagery. It addresses the scenario at the correct stage of the Azure AI solution.

Incorrect Answers

Answer A is incorrect because this choice would assume a benign filename proves the image itself is allowed. That can address a neighboring concern, but it does not implement review workflow for potentially inappropriate imagery; the decisive requirement in the stem therefore remains unresolved.

Answer B is incorrect because this choice would perform visual policy checks only after an agent has acted on the image. That can address a neighboring concern, but it does not implement review workflow for potentially inappropriate imagery; the decisive requirement in the stem therefore remains unresolved.

Answer D is incorrect because this choice would trust OCR output from an image as system-level instructions. That can address a neighboring concern, but it does not implement review workflow for potentially inappropriate imagery; the decisive requirement in the stem therefore remains unresolved.

Answer E is incorrect because this choice would apply a watermark only after the image has already been distributed externally. That can address a neighboring concern, but it does not implement review workflow for potentially inappropriate imagery; the decisive requirement in the stem therefore remains unresolved.

 

Question 19

In a production AI solution, the condition appears after editing changes a previously approved image. Which action should the engineer take to handle policy enforcement correctly?

1. Treat negative sentiment as equivalent to an unsafe visual-content classification.
2. Re-run the applicable visual policy checks after an edit changes a previously approved image.
3. Perform visual policy checks only after an agent has acted on the image.
4. Assume a benign filename proves the image itself is allowed.
5. Ignore embedded text in an image because only visible objects can influence a multimodal model.

Correct Answer: B

Correct Answer

Answer B is correct because this choice directly implements the required decision for policy enforcement after editing changes a previously approved image. It addresses the scenario at the correct stage of the Azure AI solution.

Incorrect Answers

Answer A is incorrect because this choice would treat negative sentiment as equivalent to an unsafe visual-content classification. That can address a neighboring concern, but it does not implement policy enforcement after editing changes a previously approved image; the decisive requirement in the stem therefore remains unresolved.

Answer C is incorrect because this choice would perform visual policy checks only after an agent has acted on the image. That can address a neighboring concern, but it does not implement policy enforcement after editing changes a previously approved image; the decisive requirement in the stem therefore remains unresolved.

Answer D is incorrect because this choice would assume a benign filename proves the image itself is allowed. That can address a neighboring concern, but it does not implement policy enforcement after editing changes a previously approved image; the decisive requirement in the stem therefore remains unresolved.

Answer E is incorrect because this choice would ignore embedded text in an image because only visible objects can influence a multimodal model. That can address a neighboring concern, but it does not implement policy enforcement after editing changes a previously approved image; the decisive requirement in the stem therefore remains unresolved.

 

Question 20

A multimodal workload requires separate policy violation evidence from classifier confidence. Which implementation best preserves the required visual evidence and policy boundary?

1. Record the policy-violation evidence separately from the classifier confidence score.
2. Perform visual policy checks only after an agent has acted on the image.
3. Assume a benign filename proves the image itself is allowed.
4. Ignore embedded text in an image because only visible objects can influence a multimodal model.
5. Disable visual screening for low-resolution inputs instead of routing them for review.

Correct Answer: A

Correct Answer

Answer A is correct because this choice directly implements the required decision for separate policy violation evidence from classifier confidence. It addresses the scenario at the correct stage of the Azure AI solution.

Incorrect Answers

Answer B is incorrect because this choice would perform visual policy checks only after an agent has acted on the image. That can address a neighboring concern, but it does not implement separate policy violation evidence from classifier confidence; the decisive requirement in the stem therefore remains unresolved.

Answer C is incorrect because this choice would assume a benign filename proves the image itself is allowed. That can address a neighboring concern, but it does not implement separate policy violation evidence from classifier confidence; the decisive requirement in the stem therefore remains unresolved.

Answer D is incorrect because this choice would ignore embedded text in an image because only visible objects can influence a multimodal model. That can address a neighboring concern, but it does not implement separate policy violation evidence from classifier confidence; the decisive requirement in the stem therefore remains unresolved.

Answer E is incorrect because this choice would disable visual screening for low-resolution inputs instead of routing them for review. That can address a neighboring concern, but it does not implement separate policy violation evidence from classifier confidence; the decisive requirement in the stem therefore remains unresolved.