Microsoft AI-102 (Designing and Implementing a Microsoft Azure AI Solution) Practice Test 14

Question 130

Auditors observe false blocks because common words match a prohibited substring. What should be reviewed?

  1. Distinguish content-filter outcomes from transport and service errors
  2. Use risk-based thresholds and exception rules for routing borderline or high-impact cases
  3. Evaluate the service's supported languages and test representative multilingual inputs
  4. Check the blocklist's matching behavior and the specificity of its entries

Explanation

Correct Answer

D:

Overly broad blocked terms can produce collisions; general content-severity tuning does not fix bad list design. In the public generative-AI assistant, confirm the fix by retesting the affected behavior after checking the blocklist's matching behavior and the specificity of its entries.

Incorrect Answers

A:

Policy refusals or filtered responses are different from availability failures and should trigger safe application handling in its own workflow, but that is a different correction from the present requirement to check the blocklist's matching behavior and the specificity of its entries. The reported issue would remain.

B:

Although useful elsewhere in the public generative-AI assistant, a targeted review queue focuses attention on consequences and uncertainty instead of requiring humans to approve every harmless interaction; the present case is different and calls for the team to check the blocklist's matching behavior and the specificity of its entries.

C:

That change addresses another failure mode: classifier performance and support can differ by language; a single English test set cannot establish multilingual safety; it does not address this case, for which the relevant intervention is to check the blocklist's matching behavior and the specificity of its entries.

Question 131

An AI system proposes denying a customer's financial claim. Governance requires accountable approval before notifying the customer. What flow is needed?

  1. Configure the exact organizational terms in a supported custom blocklist
  2. Queue the proposed decision for an authorized human reviewer before committing it
  3. Map supported harm-category severity results to a documented application policy
  4. Apply Prompt Shields to detect suspicious instructions embedded in retrieved content

Explanation

Correct Answer

B:

Human approval makes high-impact decisions reviewable; presenting a confidence score alone is not accountability. In the public generative-AI assistant, confirm the fix by retesting the affected behavior after queuing the proposed decision for an authorized human reviewer before committing it.

Incorrect Answers

A:

A term-specific blocklist enforces business policy independently of broad harm categories in its own workflow, but that is a different correction from the present requirement to queue the proposed decision for an authorized human reviewer before committing it. The reported issue would remain.

C:

That change addresses another failure mode: classification gives category/severity evidence; the application's policy defines the business action; it does not address this case, for which the relevant intervention is to queue the proposed decision for an authorized human reviewer before committing it.

D:

Although useful elsewhere in the public generative-AI assistant, indirect prompt injection attempts to turn lower-trust source text into instructions, distinct from ordinary malicious wording in a direct user prompt; the present case is different and calls for the team to queue the proposed decision for an authorized human reviewer before committing it.

Question 132

A team assumes every blocked output indicates a model outage. What response-handling behavior should the client add?

  1. Distinguish content-filter outcomes from transport and service errors
  2. Use risk-based thresholds and exception rules for routing borderline or high-impact cases
  3. Check the blocklist's matching behavior and the specificity of its entries
  4. Keep retrieved text at a lower trust level than system and developer instructions

Explanation

Correct Answer

A:

Policy refusals or filtered responses are different from availability failures and should trigger safe application handling. In the public generative-AI assistant, confirm the fix by retesting the affected behavior after distinguishing content-filter outcomes from transport and service errors.

Incorrect Answers

B:

Although useful elsewhere in the public generative-AI assistant, a targeted review queue focuses attention on consequences and uncertainty instead of requiring humans to approve every harmless interaction; the present case is different and calls for the team to distinguish content-filter outcomes from transport and service errors.

C:

That change addresses another failure mode: overly broad blocked terms can produce collisions; general content-severity tuning does not fix bad list design; it does not address this case, for which the relevant intervention is to distinguish content-filter outcomes from transport and service errors.

D:

External data may contain attacker instructions; retrieval relevance does not grant it authority over application behavior in its own workflow, but that is a different correction from the present requirement to distinguish content-filter outcomes from transport and service errors. The reported issue would remain.

Question 133

Prohibited terms change monthly. What maintenance procedure is needed?

  1. Minimize and secure retained text and severity metadata under an approved retention policy
  2. Add attack cases inside retrieved documents and tool responses
  3. Version and test blocklist updates before deploying them to production moderation
  4. Evaluate both harmful and benign representative prompts for false positives and misses

Explanation

Correct Answer

C:

Policy lists require controlled release practices so new terms are enforced without surprising legitimate users. In the public generative-AI assistant, confirm the fix by retesting the affected behavior after versioning and test blocklist updates before deploying them to production moderation.

Incorrect Answers

A:

Moderation telemetry may contain sensitive material; a detection use case does not justify unrestricted retention in its own workflow, but that is a different correction from the present requirement to version and test blocklist updates before deploying them to production moderation. The reported issue would remain.

B:

Although useful elsewhere in the public generative-AI assistant, indirect attacks arrive through data dependencies, not just the user's top-level message; the present case is different and calls for the team to version and test blocklist updates before deploying them to production moderation.

D:

That change addresses another failure mode: safety tuning has a utility trade-off; exclusively adversarial tests cannot show overblocking of legitimate content; it does not address this case, for which the relevant intervention is to version and test blocklist updates before deploying them to production moderation.

Question 134

An app receives a hate severity level and must decide whether to warn, block, or escalate. What should govern the action?

  1. Review the deployment's supported content-filter thresholds and policy assignments
  2. Apply Prompt Shields to detect suspicious instructions embedded in retrieved content
  3. Map supported harm-category severity results to a documented application policy
  4. Queue the proposed decision for an authorized human reviewer before committing it

Explanation

Correct Answer

C:

Classification gives category/severity evidence; the application's policy defines the business action. In the public generative-AI assistant, confirm the fix by retesting the affected behavior after maping supported harm-category severity results to a documented application policy.

Incorrect Answers

A:

Moderation behavior depends on configured content controls; changing token limits would not enforce harm-category policy in its own workflow, but that is a different correction from the present requirement to map supported harm-category severity results to a documented application policy. The reported issue would remain.

B:

That change addresses another failure mode: indirect prompt injection attempts to turn lower-trust source text into instructions, distinct from ordinary malicious wording in a direct user prompt; it does not address this case, for which the relevant intervention is to map supported harm-category severity results to a documented application policy.

D:

Although useful elsewhere in the public generative-AI assistant, human approval makes high-impact decisions reviewable; presenting a confidence score alone is not accountability; the present case is different and calls for the team to map supported harm-category severity results to a documented application policy.

Question 135

A release for a public generative-AI assistant is blocked until the application can detect attempts to override application instructions through hostile prompts or retrieved text. The tester will verify that malicious instruction attempts embedded in input are identified. What should be implemented?

  1. Evaluate harm severity with Content Safety
  2. Add a human review and escalation workflow
  3. Apply Prompt Shields
  4. Use a custom blocklist

Explanation

Correct Answer

C:

Prompt Shields evaluates suspicious direct and indirect prompt-injection patterns affecting LLM behavior. For this public generative-AI assistant, the intended evidence is that malicious instruction attempts embedded in input are identified. It does not replace account-level role assignment or prove an answer has cited sources.

Incorrect Answers

A:

Content Safety analysis returns category-specific safety assessments for text or images within service support; it may be used to classify user or generated text into supported harm categories with severity levels in a public generative-AI assistant, but the current requirement is instead to detect attempts to override application instructions through hostile prompts or retrieved text. That separate capability does not prove that malicious instruction attempts embedded in input are identified; The severity result is not an embedding similarity score or an identity-verification response, so selecting that alternative would not meet the stated acceptance check.

B:

Human-in-the-loop review provides oversight and documented exceptions where automated moderation or generation could be insufficient; it may be used to require accountable approval for high-impact or borderline AI decisions in a public generative-AI assistant, but the current requirement is instead to detect attempts to override application instructions through hostile prompts or retrieved text. That separate capability does not prove that malicious instruction attempts embedded in input are identified; A reviewer workflow cannot be replaced solely by changing temperature, so selecting that alternative would not meet the stated acceptance check.

D:

Blocklists supplement broad harm classification with explicit strings or terms configured by the application; it may be used to reject a known set of organization-specific prohibited words or phrases in a public generative-AI assistant, but the current requirement is instead to detect attempts to override application instructions through hostile prompts or retrieved text. That separate capability does not prove that malicious instruction attempts embedded in input are identified; A blocklist cannot reliably detect every paraphrase or determine whether an answer is grounded, so selecting that alternative would not meet the stated acceptance check.

Question 136

A team responsible for a public generative-AI assistant needs to restrict unsafe categories in model inputs and outputs according to deployment policy; the acceptance check is that disallowed harmful content is blocked or flagged under the configured policy. Which capability should be selected?

  1. Apply Prompt Shields
  2. Use a custom blocklist
  3. Evaluate harm severity with Content Safety
  4. Configure model content filters
  5. Add a human review and escalation workflow

Explanation

Correct Answer

D:

Content filtering assesses categories of harmful material and applies configured enforcement where supported. For this public generative-AI assistant, the intended evidence is that disallowed harmful content is blocked or flagged under the configured policy. It is not designed primarily to identify stolen API secrets or guarantee factual correctness.

Incorrect Answers

A:

Prompt Shields evaluates suspicious direct and indirect prompt-injection patterns affecting LLM behavior; it may be used to detect attempts to override application instructions through hostile prompts or retrieved text in a public generative-AI assistant, but the current requirement is instead to restrict unsafe categories in model inputs and outputs according to deployment policy. That separate capability does not prove that disallowed harmful content is blocked or flagged under the configured policy; It does not replace account-level role assignment or prove an answer has cited sources, so selecting that alternative would not meet the stated acceptance check.

B:

Blocklists supplement broad harm classification with explicit strings or terms configured by the application; it may be used to reject a known set of organization-specific prohibited words or phrases in a public generative-AI assistant, but the current requirement is instead to restrict unsafe categories in model inputs and outputs according to deployment policy. That separate capability does not prove that disallowed harmful content is blocked or flagged under the configured policy; A blocklist cannot reliably detect every paraphrase or determine whether an answer is grounded, so selecting that alternative would not meet the stated acceptance check.

C:

Content Safety analysis returns category-specific safety assessments for text or images within service support; it may be used to classify user or generated text into supported harm categories with severity levels in a public generative-AI assistant, but the current requirement is instead to restrict unsafe categories in model inputs and outputs according to deployment policy. That separate capability does not prove that disallowed harmful content is blocked or flagged under the configured policy; The severity result is not an embedding similarity score or an identity-verification response, so selecting that alternative would not meet the stated acceptance check.

E:

Human-in-the-loop review provides oversight and documented exceptions where automated moderation or generation could be insufficient; it may be used to require accountable approval for high-impact or borderline AI decisions in a public generative-AI assistant, but the current requirement is instead to restrict unsafe categories in model inputs and outputs according to deployment policy. That separate capability does not prove that disallowed harmful content is blocked or flagged under the configured policy; A reviewer workflow cannot be replaced solely by changing temperature, so selecting that alternative would not meet the stated acceptance check.

Question 137

The solution brief for a public generative-AI assistant requires the system to require accountable approval for high-impact or borderline AI decisions. A successful test confirms that sensitive outputs are held for review before downstream action. What should the team use?

  1. Use a custom blocklist
  2. Add a human review and escalation workflow
  3. Evaluate harm severity with Content Safety
  4. Configure model content filters

Explanation

Correct Answer

B:

Human-in-the-loop review provides oversight and documented exceptions where automated moderation or generation could be insufficient. For this public generative-AI assistant, the intended evidence is that sensitive outputs are held for review before downstream action. A reviewer workflow cannot be replaced solely by changing temperature.

Incorrect Answers

A:

Blocklists supplement broad harm classification with explicit strings or terms configured by the application; it may be used to reject a known set of organization-specific prohibited words or phrases in a public generative-AI assistant, but the current requirement is instead to require accountable approval for high-impact or borderline AI decisions. That separate capability does not prove that sensitive outputs are held for review before downstream action; A blocklist cannot reliably detect every paraphrase or determine whether an answer is grounded, so selecting that alternative would not meet the stated acceptance check.

C:

Content Safety analysis returns category-specific safety assessments for text or images within service support; it may be used to classify user or generated text into supported harm categories with severity levels in a public generative-AI assistant, but the current requirement is instead to require accountable approval for high-impact or borderline AI decisions. That separate capability does not prove that sensitive outputs are held for review before downstream action; The severity result is not an embedding similarity score or an identity-verification response, so selecting that alternative would not meet the stated acceptance check.

D:

Content filtering assesses categories of harmful material and applies configured enforcement where supported; it may be used to restrict unsafe categories in model inputs and outputs according to deployment policy in a public generative-AI assistant, but the current requirement is instead to require accountable approval for high-impact or borderline AI decisions. That separate capability does not prove that sensitive outputs are held for review before downstream action; It is not designed primarily to identify stolen API secrets or guarantee factual correctness, so selecting that alternative would not meet the stated acceptance check.

Question 138

A team stores all user prompts indefinitely to investigate moderation scores. What privacy issue should be assessed?

  1. Minimize and secure retained text and severity metadata under an approved retention policy
  2. Version and test blocklist updates before deploying them to production moderation
  3. Evaluate both harmful and benign representative prompts for false positives and misses
  4. Add attack cases inside retrieved documents and tool responses

Explanation

Correct Answer

A:

Moderation telemetry may contain sensitive material; a detection use case does not justify unrestricted retention. In the public generative-AI assistant, confirm the fix by retesting the affected behavior after minimizing and secure retained text and severity metadata under an approved retention policy.

Incorrect Answers

B:

Policy lists require controlled release practices so new terms are enforced without surprising legitimate users in its own workflow, but that is a different correction from the present requirement to minimize and secure retained text and severity metadata under an approved retention policy. The reported issue would remain.

C:

That change addresses another failure mode: safety tuning has a utility trade-off; exclusively adversarial tests cannot show overblocking of legitimate content; it does not address this case, for which the relevant intervention is to minimize and secure retained text and severity metadata under an approved retention policy.

D:

Although useful elsewhere in the public generative-AI assistant, indirect attacks arrive through data dependencies, not just the user's top-level message; the present case is different and calls for the team to minimize and secure retained text and severity metadata under an approved retention policy.

Question 139

Reviewers are overloaded with low-risk chatbot greetings while dangerous cases are missed. How should escalation be improved?

  1. Keep retrieved text at a lower trust level than system and developer instructions
  2. Use risk-based thresholds and exception rules for routing borderline or high-impact cases
  3. Evaluate the service's supported languages and test representative multilingual inputs
  4. Distinguish content-filter outcomes from transport and service errors

Explanation

Correct Answer

B:

A targeted review queue focuses attention on consequences and uncertainty instead of requiring humans to approve every harmless interaction. In the public generative-AI assistant, confirm the fix by retesting the affected behavior after using risk-based thresholds and exception rules for routing borderline or high-impact cases.

Incorrect Answers

A:

That change addresses another failure mode: external data may contain attacker instructions; retrieval relevance does not grant it authority over application behavior; it does not address this case, for which the relevant intervention is to use risk-based thresholds and exception rules for routing borderline or high-impact cases.

C:

Classifier performance and support can differ by language; a single English test set cannot establish multilingual safety in its own workflow, but that is a different correction from the present requirement to use risk-based thresholds and exception rules for routing borderline or high-impact cases. The reported issue would remain.

D:

Although useful elsewhere in the public generative-AI assistant, policy refusals or filtered responses are different from availability failures and should trigger safe application handling; the present case is different and calls for the team to use risk-based thresholds and exception rules for routing borderline or high-impact cases.

Question 140

A company must block a short list of prohibited product names, even if a general harm classifier rates the sentence harmless. Which control is appropriate?

  1. Review the deployment's supported content-filter thresholds and policy assignments
  2. Configure the exact organizational terms in a supported custom blocklist
  3. Queue the proposed decision for an authorized human reviewer before committing it
  4. Apply Prompt Shields to detect suspicious instructions embedded in retrieved content

Explanation

Correct Answer

B:

A term-specific blocklist enforces business policy independently of broad harm categories. In the public generative-AI assistant, confirm the fix by retesting the affected behavior after configuring the exact organizational terms in a supported custom blocklist.

Incorrect Answers

A:

Moderation behavior depends on configured content controls; changing token limits would not enforce harm-category policy in its own workflow, but that is a different correction from the present requirement to configure the exact organizational terms in a supported custom blocklist. The reported issue would remain.

C:

That change addresses another failure mode: human approval makes high-impact decisions reviewable; presenting a confidence score alone is not accountability; it does not address this case, for which the relevant intervention is to configure the exact organizational terms in a supported custom blocklist.

D:

Although useful elsewhere in the public generative-AI assistant, indirect prompt injection attempts to turn lower-trust source text into instructions, distinct from ordinary malicious wording in a direct user prompt; the present case is different and calls for the team to configure the exact organizational terms in a supported custom blocklist.

Designing and Implementing a Microsoft Azure AI Solution (AI-102) Practice Tests