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 43

Question 420

During a cost review, a quality assurance lead must accept a change for output safety after retrieval. The issue to be addressed is: A RAG answer reproduces a secret visible in retrieved system-maintenance notes. Which concrete verification action demonstrates that the relevant behavior works?

  1. Ask a known conflicting-policy question and inspect the selected effective version
  2. Trust a final wording filter to secure a globally privileged retriever
  3. Query for a seeded secret using an unprivileged identity and confirm non-disclosure
  4. Redact or exclude sensitive source passages before exposing them to the generation step

Explanation

Correct Answer

C:

To verify this change, the legal operations unit should query for a seeded secret using an unprivileged identity and confirm non-disclosure. Operators still need the implementation to redact or exclude sensitive source passages before exposing them to the generation step and the preventive requirement to require sensitive-data scans and per-user access enforcement.

Incorrect Answers

A:

For cross-document synthesis, it can be sensible to ask a known conflicting-policy question and inspect the selected effective version if a compliance answer must reconcile two approved documents with different effective dates. Without the actual change to query for a seeded secret using an unprivileged identity and confirm non-disclosure, the answer lacks the correct operational scope even if the alternate action is technically valid elsewhere.

B:

In the legal operations unit acceptance review, trying to trust a final wording filter to secure a globally privileged retriever does not test the exact output safety after retrieval condition where a RAG answer reproduces a secret visible in retrieved system-maintenance notes; the relevant test is to query for a seeded secret using an unprivileged identity and confirm non-disclosure. Acceptance should instead demonstrate that the service can query for a seeded secret using an unprivileged identity and confirm non-disclosure, using the described failure as the test criterion.

D:

The action to redact or exclude sensitive source passages before exposing them to the generation step can contribute to output safety after retrieval as a implementation step; here, a RAG answer reproduces a secret visible in retrieved system-maintenance notes. In the legal operations unit acceptance review, trying to redact or exclude sensitive source passages before exposing them to the generation step does not test the exact output safety after retrieval condition where a RAG answer reproduces a secret visible in retrieved system-maintenance notes; the relevant test is to query for a seeded secret using an unprivileged identity and confirm non-disclosure.

Question 421

At a scientific publisher, the AI platform engineer is preparing a support escalation. The concrete requirement is: A new retrieval prompt sounds fluent but invents technical product restrictions. Which design action BEST addresses the stated need?

  1. Deploy a compatible Foundry conversational model and connect the application to its endpoint
  2. Evaluate factual groundedness using source-supported queries and adversarial missing-fact cases
  3. Declare a prompt improved because its answers became longer and more confident
  4. Run supported and unsupported test questions and score factual support

Explanation

Correct Answer

B:

The proposed implementation is to evaluate factual groundedness using source-supported queries and adversarial missing-fact cases; this matches the operational requirement that a new retrieval prompt sounds fluent but invents technical product restrictions. The team can run supported and unsupported test questions and score factual support. The team should also maintain versioned factual evaluation examples linked to original sources.

Incorrect Answers

A:

Choosing to deploy a compatible Foundry conversational model and connect the application to its endpoint would direct effort toward model deployment for chat, whose characteristic issue is that a customer portal has no model endpoint capable of supporting conversational completion. The relevant corrective or preventive action here remains to evaluate factual groundedness using source-supported queries and adversarial missing-fact cases, with its own validation criteria.

C:

Choosing to declare a prompt improved because its answers became longer and more confident cannot directly replace the specified action to evaluate factual groundedness using source-supported queries and adversarial missing-fact cases when a new retrieval prompt sounds fluent but invents technical product restrictions. The actual implementation must evaluate factual groundedness using source-supported queries and adversarial missing-fact cases, because the stated service condition is that a new retrieval prompt sounds fluent but invents technical product restrictions.

D:

The action to run supported and unsupported test questions and score factual support can contribute to model evaluation for fabrication as a targeted acceptance check step; here, a new retrieval prompt sounds fluent but invents technical product restrictions. Choosing to run supported and unsupported test questions and score factual support cannot directly replace the specified action to evaluate factual groundedness using source-supported queries and adversarial missing-fact cases when a new retrieval prompt sounds fluent but invents technical product restrictions.

Question 422

After a service audit at a nonprofit organization, the team observes this failure: A new retrieval prompt sounds fluent but invents technical product restrictions. Which flawed assumption or implementation choice would MOST DIRECTLY explain why the intended outcome was not achieved?

  1. Run supported and unsupported test questions and score factual support
  2. Declare a prompt improved because its answers became longer and more confident
  3. Assume provisioning a Foundry project automatically deploys every model
  4. Evaluate factual groundedness using source-supported queries and adversarial missing-fact cases

Explanation

Correct Answer

B:

This design mistake is to declare a prompt improved because its answers became longer and more confident; it conflicts with the actual model evaluation for fabrication behavior needed when a new retrieval prompt sounds fluent but invents technical product restrictions. The nonprofit organization should instead evaluate factual groundedness using source-supported queries and adversarial missing-fact cases, followed by a check to run supported and unsupported test questions and score factual support.

Incorrect Answers

A:

The alternative to run supported and unsupported test questions and score factual support is a targeted acceptance check step rather than the faulty assumption being diagnosed. This nonprofit organization failure-analysis alternative of run supported and unsupported test questions and score factual support does not diagnose the particular model evaluation for fabrication defect where a new retrieval prompt sounds fluent but invents technical product restrictions; the keyed mistake is declare a prompt improved because its answers became longer and more confident.

C:

The option to assume provisioning a Foundry project automatically deploys every model would normally serve a different model deployment for chat requirement, where a customer portal has no model endpoint capable of supporting conversational completion. The alternative may be useful in its own context, but it does not directly supply declare a prompt improved because its answers became longer and more confident for the stated decision.

D:

The alternative to evaluate factual groundedness using source-supported queries and adversarial missing-fact cases is a implementation step rather than the faulty assumption being diagnosed. This nonprofit organization failure-analysis alternative of evaluate factual groundedness using source-supported queries and adversarial missing-fact cases does not diagnose the particular model evaluation for fabrication defect where a new retrieval prompt sounds fluent but invents technical product restrictions; the keyed mistake is declare a prompt improved because its answers became longer and more confident.

Question 423

The municipal services team is responding to a security tabletop exercise. A product support lead discovers this issue: A new retrieval prompt sounds fluent but invents technical product restrictions. What corrective step should the team take to restore the intended functionality most directly?

  1. Restore the prior prompt or evidence constraints after demonstrated regression
  2. Run supported and unsupported test questions and score factual support
  3. Declare a prompt improved because its answers became longer and more confident
  4. Create or repair the model deployment and update application endpoint configuration

Explanation

Correct Answer

A:

The direct repair is to restore the prior prompt or evidence constraints after demonstrated regression. Once the condition is corrected, the municipal services team should run supported and unsupported test questions and score factual support. To reduce recurrence, it should also maintain versioned factual evaluation examples linked to original sources; those steps provide different recovery and prevention functions.

Incorrect Answers

B:

The action to run supported and unsupported test questions and score factual support can contribute to model evaluation for fabrication as a targeted acceptance check step; here, a new retrieval prompt sounds fluent but invents technical product restrictions. During municipal services team remediation, the proposed step to run supported and unsupported test questions and score factual support is not the direct restoration action for model evaluation for fabrication when a new retrieval prompt sounds fluent but invents technical product restrictions; the keyed repair is restore the prior prompt or evidence constraints after demonstrated regression.

C:

During municipal services team remediation, the proposed step to declare a prompt improved because its answers became longer and more confident is not the direct restoration action for model evaluation for fabrication when a new retrieval prompt sounds fluent but invents technical product restrictions; the keyed repair is restore the prior prompt or evidence constraints after demonstrated regression. After remediation the team must run supported and unsupported test questions and score factual support; the action in this option does not itself establish that target recovery outcome.

D:

This approach relates to model deployment for chat: a customer portal has no model endpoint capable of supporting conversational completion. If the team chose it instead, the targeted check to run supported and unsupported test questions and score factual support could still fail because the underlying required action is to restore the prior prompt or evidence constraints after demonstrated regression.

Question 424

A data engineer at a payment services team wants the following issue to be prevented in later deployments: A new retrieval prompt sounds fluent but invents technical product restrictions. Which enduring control or engineering guardrail BEST prevents recurrence?

  1. Maintain versioned factual evaluation examples linked to original sources
  2. Require visual evidence citations and accessible descriptions
  3. Restore the prior prompt or evidence constraints after demonstrated regression
  4. Declare a prompt improved because its answers became longer and more confident

Explanation

Correct Answer

A:

Standardizing the boundary to maintain versioned factual evaluation examples linked to original sources is preventive because the same model evaluation for fabrication failure could otherwise return: a new retrieval prompt sounds fluent but invents technical product restrictions. For this payment services team, the service should also evaluate factual groundedness using source-supported queries and adversarial missing-fact cases, and tests should run supported and unsupported test questions and score factual support so that the ongoing requirement is measurable.

Incorrect Answers

B:

The action to require visual evidence citations and accessible descriptions addresses multimodal document question answering rather than the decisive model evaluation for fabrication condition. The payment services team should not infer that fixing one adjacent AI service validates its unrelated workflow; the requested action is to maintain versioned factual evaluation examples linked to original sources.

C:

The action to restore the prior prompt or evidence constraints after demonstrated regression can contribute to model evaluation for fabrication as a corrective remediation step; here, a new retrieval prompt sounds fluent but invents technical product restrictions. For prevention at the payment services team, the measure to restore the prior prompt or evidence constraints after demonstrated regression is weaker or aimed at another issue than the model evaluation for fabrication guardrail to maintain versioned factual evaluation examples linked to original sources needed when a new retrieval prompt sounds fluent but invents technical product restrictions.

D:

For prevention at the payment services team, the measure to declare a prompt improved because its answers became longer and more confident is weaker or aimed at another issue than the model evaluation for fabrication guardrail to maintain versioned factual evaluation examples linked to original sources needed when a new retrieval prompt sounds fluent but invents technical product restrictions. The lasting guardrail for this specific threat is to maintain versioned factual evaluation examples linked to original sources, not simply to complete an adjacent service task.

Question 425

During a service audit, a service administrator must accept a change for model evaluation for fabrication. The issue to be addressed is: A new retrieval prompt sounds fluent but invents technical product restrictions. Which concrete verification action demonstrates that the relevant behavior works?

  1. Declare a prompt improved because its answers became longer and more confident
  2. Maintain versioned factual evaluation examples linked to original sources
  3. Restore the prior prompt or evidence constraints after demonstrated regression
  4. Retrieve an exact part number and a paraphrased fault description
  5. Evaluate factual groundedness using source-supported queries and adversarial missing-fact cases
  6. Run supported and unsupported test questions and score factual support

Explanation

Correct Answer

F:

The targeted acceptance test is to run supported and unsupported test questions and score factual support. The design should first evaluate factual groundedness using source-supported queries and adversarial missing-fact cases. A robust transport authority deployment also needs to maintain versioned factual evaluation examples linked to original sources; passing this check provides relevant evidence but does not certify every possible input or security risk.

Incorrect Answers

A:

In the transport authority acceptance review, trying to declare a prompt improved because its answers became longer and more confident does not test the exact model evaluation for fabrication condition where a new retrieval prompt sounds fluent but invents technical product restrictions; the relevant test is to run supported and unsupported test questions and score factual support. Acceptance should instead demonstrate that the service can run supported and unsupported test questions and score factual support, using the described failure as the test criterion.

B:

The action to maintain versioned factual evaluation examples linked to original sources can contribute to model evaluation for fabrication as a ongoing preventive rule step; here, a new retrieval prompt sounds fluent but invents technical product restrictions. In the transport authority acceptance review, trying to maintain versioned factual evaluation examples linked to original sources does not test the exact model evaluation for fabrication condition where a new retrieval prompt sounds fluent but invents technical product restrictions; the relevant test is to run supported and unsupported test questions and score factual support.

C:

The action to restore the prior prompt or evidence constraints after demonstrated regression can contribute to model evaluation for fabrication as a corrective remediation step; here, a new retrieval prompt sounds fluent but invents technical product restrictions. In the transport authority acceptance review, trying to restore the prior prompt or evidence constraints after demonstrated regression does not test the exact model evaluation for fabrication condition where a new retrieval prompt sounds fluent but invents technical product restrictions; the relevant test is to run supported and unsupported test questions and score factual support.

D:

The option to retrieve an exact part number and a paraphrased fault description would normally serve a different hybrid retrieval for identifiers requirement, where users search for part numbers and semantically related troubleshooting symptoms. The alternative may be useful in its own context, but it does not directly supply run supported and unsupported test questions and score factual support for the stated decision.

E:

The action to evaluate factual groundedness using source-supported queries and adversarial missing-fact cases can contribute to model evaluation for fabrication as a implementation step; here, a new retrieval prompt sounds fluent but invents technical product restrictions. In the transport authority acceptance review, trying to evaluate factual groundedness using source-supported queries and adversarial missing-fact cases does not test the exact model evaluation for fabrication condition where a new retrieval prompt sounds fluent but invents technical product restrictions; the relevant test is to run supported and unsupported test questions and score factual support.

Question 426

At a legal operations unit, the cloud developer is preparing a release acceptance review. The concrete requirement is: An agent attempts unrelated tasks because its service role is poorly scoped. Which design action BEST addresses the stated need?

  1. Ask for an unrelated privileged operation and confirm refusal or escalation
  2. Define narrow objectives allowed actions and stop conditions in agent configuration
  3. Assume a broad instruction to help users safely limits tool behavior
  4. Return source metadata tied to each important retrieved claim

Explanation

Correct Answer

B:

Choosing to define narrow objectives allowed actions and stop conditions in agent configuration directly targets the described agent goal definition gap: an agent attempts unrelated tasks because its service role is poorly scoped. As a lasting boundary, they also need to require explicit authorized goal and prohibited-action tests, because an implementation by itself does not prove safe operation.

Incorrect Answers

A:

The action to ask for an unrelated privileged operation and confirm refusal or escalation can contribute to agent goal definition as a targeted acceptance check step; here, an agent attempts unrelated tasks because its service role is poorly scoped. Choosing to ask for an unrelated privileged operation and confirm refusal or escalation cannot directly replace the specified action to define narrow objectives allowed actions and stop conditions in agent configuration when an agent attempts unrelated tasks because its service role is poorly scoped.

C:

Choosing to assume a broad instruction to help users safely limits tool behavior cannot directly replace the specified action to define narrow objectives allowed actions and stop conditions in agent configuration when an agent attempts unrelated tasks because its service role is poorly scoped. The actual implementation must define narrow objectives allowed actions and stop conditions in agent configuration, because the stated service condition is that an agent attempts unrelated tasks because its service role is poorly scoped.

D:

Choosing to return source metadata tied to each important retrieved claim cannot directly replace the specified action to define narrow objectives allowed actions and stop conditions in agent configuration when an agent attempts unrelated tasks because its service role is poorly scoped. The relevant corrective or preventive action here remains to define narrow objectives allowed actions and stop conditions in agent configuration, with its own validation criteria.

Question 427

After a design meeting at a financial services division, the team observes this failure: An agent attempts unrelated tasks because its service role is poorly scoped. Which flawed assumption or implementation choice would MOST DIRECTLY explain why the intended outcome was not achieved?

  1. Ask for an unrelated privileged operation and confirm refusal or escalation
  2. Return unstructured prose and assume downstream automation will parse it correctly
  3. Assume a broad instruction to help users safely limits tool behavior
  4. Define narrow objectives allowed actions and stop conditions in agent configuration

Explanation

Correct Answer

C:

The faulty assumption is assume a broad instruction to help users safely limits tool behavior. A functioning alternative would be to define narrow objectives allowed actions and stop conditions in agent configuration, and the relevant validation would be to ask for an unrelated privileged operation and confirm refusal or escalation.

Incorrect Answers

A:

The alternative to ask for an unrelated privileged operation and confirm refusal or escalation is a targeted acceptance check step rather than the faulty assumption being diagnosed. This financial services division failure-analysis alternative of ask for an unrelated privileged operation and confirm refusal or escalation does not diagnose the particular agent goal definition defect where an agent attempts unrelated tasks because its service role is poorly scoped; the keyed mistake is assume a broad instruction to help users safely limits tool behavior.

B:

This financial services division failure-analysis alternative of return unstructured prose and assume downstream automation will parse it correctly does not diagnose the particular agent goal definition defect where an agent attempts unrelated tasks because its service role is poorly scoped; the keyed mistake is assume a broad instruction to help users safely limits tool behavior. The alternative may be useful in its own context, but it does not directly supply assume a broad instruction to help users safely limits tool behavior for the stated decision.

D:

The alternative to define narrow objectives allowed actions and stop conditions in agent configuration is a implementation step rather than the faulty assumption being diagnosed. This financial services division failure-analysis alternative of define narrow objectives allowed actions and stop conditions in agent configuration does not diagnose the particular agent goal definition defect where an agent attempts unrelated tasks because its service role is poorly scoped; the keyed mistake is assume a broad instruction to help users safely limits tool behavior.

Question 428

The logistics operator is responding to a staged service rollout. A security engineer discovers this issue: An agent attempts unrelated tasks because its service role is poorly scoped. What corrective step should the team take to restore the intended functionality most directly?

  1. Narrow agent instructions and tool permissions then replay the unsafe workflow
  2. Ask for an unrelated privileged operation and confirm refusal or escalation
  3. Rebuild state from a verified checkpoint and resume only unfinished steps
  4. Assume a broad instruction to help users safely limits tool behavior

Explanation

Correct Answer

A:

The appropriate remediation is to narrow agent instructions and tool permissions then replay the unsafe workflow, because the failure concerns agent goal definition: an agent attempts unrelated tasks because its service role is poorly scoped. A meaningful verification should ask for an unrelated privileged operation and confirm refusal or escalation, and the durable operating rule should require explicit authorized goal and prohibited-action tests after the immediate fix.

Incorrect Answers

B:

The action to ask for an unrelated privileged operation and confirm refusal or escalation can contribute to agent goal definition as a targeted acceptance check step; here, an agent attempts unrelated tasks because its service role is poorly scoped. During logistics operator remediation, the proposed step to ask for an unrelated privileged operation and confirm refusal or escalation is not the direct restoration action for agent goal definition when an agent attempts unrelated tasks because its service role is poorly scoped; the keyed repair is narrow agent instructions and tool permissions then replay the unsafe workflow.

C:

This approach relates to agent context compaction: a long-running agent exhausts its context and forgets constraints. If the team chose it instead, the targeted check to ask for an unrelated privileged operation and confirm refusal or escalation could still fail because the underlying required action is to narrow agent instructions and tool permissions then replay the unsafe workflow.

D:

During logistics operator remediation, the proposed step to assume a broad instruction to help users safely limits tool behavior is not the direct restoration action for agent goal definition when an agent attempts unrelated tasks because its service role is poorly scoped; the keyed repair is narrow agent instructions and tool permissions then replay the unsafe workflow. After remediation the team must ask for an unrelated privileged operation and confirm refusal or escalation; the action in this option does not itself establish that target recovery outcome.

Question 429

A agent designer at a energy cooperative wants the following issue to be prevented in later deployments: An agent attempts unrelated tasks because its service role is poorly scoped. Which enduring control or engineering guardrail BEST prevents recurrence?

  1. Require retention schedules and access audit for memory
  2. Narrow agent instructions and tool permissions then replay the unsafe workflow
  3. Require explicit authorized goal and prohibited-action tests
  4. Ask for an unrelated privileged operation and confirm refusal or escalation
  5. Assume a broad instruction to help users safely limits tool behavior

Explanation

Correct Answer

C:

The sustained prevention measure is to require explicit authorized goal and prohibited-action tests. The implementation can still use the separate action to define narrow objectives allowed actions and stop conditions in agent configuration, but a guardrail constrains future changes. The energy cooperative can measure whether the rule works by arranging to ask for an unrelated privileged operation and confirm refusal or escalation.

Incorrect Answers

A:

For memory retention for agents, it can be sensible to require retention schedules and access audit for memory if an agent stores user conversations indefinitely despite retention obligations. Without the actual change to require explicit authorized goal and prohibited-action tests, the answer lacks the correct operational scope even if the alternate action is technically valid elsewhere.

B:

The action to narrow agent instructions and tool permissions then replay the unsafe workflow can contribute to agent goal definition as a corrective remediation step; here, an agent attempts unrelated tasks because its service role is poorly scoped. For prevention at the energy cooperative, the measure to narrow agent instructions and tool permissions then replay the unsafe workflow is weaker or aimed at another issue than the agent goal definition guardrail to require explicit authorized goal and prohibited-action tests needed when an agent attempts unrelated tasks because its service role is poorly scoped.

D:

The action to ask for an unrelated privileged operation and confirm refusal or escalation can contribute to agent goal definition as a targeted acceptance check step; here, an agent attempts unrelated tasks because its service role is poorly scoped. For prevention at the energy cooperative, the measure to ask for an unrelated privileged operation and confirm refusal or escalation is weaker or aimed at another issue than the agent goal definition guardrail to require explicit authorized goal and prohibited-action tests needed when an agent attempts unrelated tasks because its service role is poorly scoped.

E:

For prevention at the energy cooperative, the measure to assume a broad instruction to help users safely limits tool behavior is weaker or aimed at another issue than the agent goal definition guardrail to require explicit authorized goal and prohibited-action tests needed when an agent attempts unrelated tasks because its service role is poorly scoped. The lasting guardrail for this specific threat is to require explicit authorized goal and prohibited-action tests, not simply to complete an adjacent service task.

Question 430

During a design meeting, a responsible-AI reviewer must accept a change for agent goal definition. The issue to be addressed is: An agent attempts unrelated tasks because its service role is poorly scoped. Which concrete verification action demonstrates that the relevant behavior works?

  1. Ask for an unrelated privileged operation and confirm refusal or escalation
  2. Submit a missing-authorization handoff and confirm it is rejected
  3. Assume a broad instruction to help users safely limits tool behavior
  4. Define narrow objectives allowed actions and stop conditions in agent configuration

Explanation

Correct Answer

A:

To verify this change, the industrial maintenance team should ask for an unrelated privileged operation and confirm refusal or escalation. Operators still need the implementation to define narrow objectives allowed actions and stop conditions in agent configuration and the preventive requirement to require explicit authorized goal and prohibited-action tests.

Incorrect Answers

B:

For multi-agent handoff schema, it can be sensible to submit a missing-authorization handoff and confirm it is rejected if a planner agent passes incomplete instructions to an executor agent. Without the actual change to ask for an unrelated privileged operation and confirm refusal or escalation, the answer lacks the correct operational scope even if the alternate action is technically valid elsewhere.

C:

In the industrial maintenance team acceptance review, trying to assume a broad instruction to help users safely limits tool behavior does not test the exact agent goal definition condition where an agent attempts unrelated tasks because its service role is poorly scoped; the relevant test is to ask for an unrelated privileged operation and confirm refusal or escalation. Acceptance should instead demonstrate that the service can ask for an unrelated privileged operation and confirm refusal or escalation, using the described failure as the test criterion.

D:

The action to define narrow objectives allowed actions and stop conditions in agent configuration can contribute to agent goal definition as a implementation step; here, an agent attempts unrelated tasks because its service role is poorly scoped. In the industrial maintenance team acceptance review, trying to define narrow objectives allowed actions and stop conditions in agent configuration does not test the exact agent goal definition condition where an agent attempts unrelated tasks because its service role is poorly scoped; the relevant test is to ask for an unrelated privileged operation and confirm refusal or escalation.

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