Pass Microsoft AI-103 Exam in First Attempt Easily
Real Microsoft AI-103 Exam Questions, Accurate & Verified Answers As Experienced in the Actual Test!

Verified by experts

AI-103 Premium File

  • 187 Questions & Answers
  • Last Update: Nov 19, 2026
$69.99 $76.99

Developing AI Apps and Agents on Azure AI-103 Practice Test 19

Question 180

An Azure AI engineer must address the following production request: An agent reads an online knowledge page that can contain attacker-controlled text. The solution must keep retrieved instructions separate from privileged system and tool policy. What should be implemented?

  1. Mark retrieved content as untrusted, limit tool rights, and enforce authorization outside the model.
  2. Automatically call tools named in retrieved Markdown without validation.
  3. Skip all tool logging when a page looks authoritative.
  4. Permit page text to redefine the agent's system role.

Explanation

Correct Answer

A:

A web page cannot be allowed to order privileged tool calls; for this workload the decision is to mark retrieved content as untrusted, limit tool rights, and enforce authorization outside the model. The acceptance result is: Inject a mock exfiltration instruction in a test page and observe tool-call denial. An independent edge-case check is: Confirm the agent still answers legitimate questions based on page facts. That distinction matters because even robust prompt defenses require network and secret-access controls.

Incorrect Answers

B:

The page may contain malicious commands, which does not satisfy the requirement to keep retrieved instructions separate from privileged system and tool policy. A suitable test of the correct implementation is: Inject a mock exfiltration instruction in a test page and observe tool-call denial. This catches the particular incident where reading a page causes an unexpected outbound transfer of internal data.

C:

Source appearance does not prove safe execution. This shortcut can leave the following observable failure unresolved: Reading a page causes an unexpected outbound transfer of internal data. A proper deployment instead must keep retrieved instructions separate from privileged system and tool policy; the missing capability would leave the deployment exposed to reading a page causes an unexpected outbound transfer of internal data. The right approach is to mark retrieved content as untrusted, limit tool rights, and enforce authorization outside the model, rather than relying on this unrelated shortcut. The distinguishing test is: Inject a mock exfiltration instruction in a test page and observe tool-call denial.

D:

External content must not outrank developer instructions, so it cannot establish keep retrieved instructions separate from privileged system and tool policy. The working design instead needs to mark retrieved content as untrusted, limit tool rights, and enforce authorization outside the model. A strong boundary test is: Confirm the agent still answers legitimate questions based on page facts. This matters because even robust prompt defenses require network and secret-access controls.

Question 181

A Foundry application has a specific incident: Reading a page causes an unexpected outbound transfer of internal data. The operator is trying to keep retrieved instructions separate from privileged system and tool policy. Which failure mechanism best explains the evidence?

  1. The agent retried a timed-out write without idempotency or reconciliation.
  2. Tool argument mapping failed to enforce required typed schema.
  3. The orchestrator failed to preserve source provenance and role-specific outputs.
  4. The orchestration pipeline treated retrieved page instructions as executable authority.

Explanation

Correct Answer

D:

The failure that reading a page causes an unexpected outbound transfer of internal data points to the orchestration pipeline treated retrieved page instructions as executable authority. The direct functional check is: Inject a mock exfiltration instruction in a test page and observe tool-call denial. Then test the independent boundary—confirm the agent still answers legitimate questions based on page facts—to distinguish a genuine repair from a passing routine response.

Incorrect Answers

A:

If the agent retried a timed-out write without idempotency or reconciliation, then the expected symptom is two reservations appear for the same trip after one user request; this incident instead reports that reading a page causes an unexpected outbound transfer of internal data. The nearer diagnosis is the orchestration pipeline treated retrieved page instructions as executable authority; to distinguish it from the alternative of the agent retried a timed-out write without idempotency or reconciliation, test the specific service behavior: Inject a mock exfiltration instruction in a test page and observe tool-call denial.

B:

If tool argument mapping failed to enforce required typed schema, then the expected symptom is a locale field contains a product name and the catalog returns unrelated content; this incident instead reports that reading a page causes an unexpected outbound transfer of internal data. The nearer diagnosis is the orchestration pipeline treated retrieved page instructions as executable authority; to distinguish it from the alternative of tool argument mapping failed to enforce required typed schema, test the specific service behavior: Inject a mock exfiltration instruction in a test page and observe tool-call denial.

C:

If the orchestrator failed to preserve source provenance and role-specific outputs, then the expected symptom is the final answer cites a translation agent for a claim that required retrieval; this incident instead reports that reading a page causes an unexpected outbound transfer of internal data. The nearer diagnosis is the orchestration pipeline treated retrieved page instructions as executable authority; to distinguish it from the alternative of the orchestrator failed to preserve source provenance and role-specific outputs, test the specific service behavior: Inject a mock exfiltration instruction in a test page and observe tool-call denial.

Question 182

A reviewer must verify that the solution will keep retrieved instructions separate from privileged system and tool policy. The deployment context is: An agent reads an online knowledge page that can contain attacker-controlled text. Which TWO tests provide the strongest evidence? Select TWO answers.

  1. Block every webpage tool call after detecting injected text and claim the agent remains useful.
  2. Confirm the agent still answers legitimate questions based on page facts.
  3. Verify the injection string appears in a safety report but do not check whether a data-transfer tool was invoked.
  4. Inject a mock exfiltration instruction in a test page and observe tool-call denial.
  5. Trust an instruction because it appears in a fetched page's bold text rather than the user request.

Explanation

Correct Answers

B:

Confirm the agent still answers legitimate questions based on page facts exercises a relevant negative or edge-case condition. It matters here because even robust prompt defenses require network and secret-access controls. The complementary positive check—inject a mock exfiltration instruction in a test page and observe tool-call denial—then confirms the intended business capability still works; neither observation alone settles the whole requirement.

D:

Inject a mock exfiltration instruction in a test page and observe tool-call denial measures the actual outcome needed to keep retrieved instructions separate from privileged system and tool policy. It would expose the specific failure where reading a page causes an unexpected outbound transfer of internal data. Pair it with confirm the agent still answers legitimate questions based on page facts to check both successful behavior and the remaining operational boundary.

Incorrect Answers

A:

A full block can prevent benign page-grounded questions from being answered. For this workload, the decisive functional evidence is: Inject a mock exfiltration instruction in a test page and observe tool-call denial. The shortcut in this option could pass even when reading a page causes an unexpected outbound transfer of internal data.

C:

Detection logs alone cannot prove the unauthorized side effect was denied. The deployment needs to keep retrieved instructions separate from privileged system and tool policy. An independent test is: Confirm the agent still answers legitimate questions based on page facts.

E:

Page content is lower-trust evidence, not authority to send private data. The real incident to guard against is that reading a page causes an unexpected outbound transfer of internal data. This does not replace the case-specific validation: Confirm the agent still answers legitimate questions based on page facts.

Question 183

Consider an application deployed with Microsoft Foundry: A field-service agent must call a maintenance API that returns tool-specific errors such as missing asset or forbidden location. The owner wants to map external errors into bounded user-visible outcomes. Which implementation is most appropriate?

  1. Handle status codes distinctly, preserve correlation IDs, and never invent success after failure.
  2. Hide tool error details entirely and claim the asset is repaired.
  3. Convert every non-200 response into a successful generic recommendation.
  4. Retry forbidden requests by switching service accounts.

Explanation

Correct Answer

A:

Operational errors need accurate reporting and safe retry decisions; for this workload the decision is to handle status codes distinctly, preserve correlation IDs, and never invent success after failure. The acceptance result is: Test forbidden and not-found responses and inspect user-facing status and traces. An independent edge-case check is: Recover from a transient 503 without duplicating maintenance tickets. That distinction matters because different failures have different safe retry rules and operational consequences.

Incorrect Answers

B:

False success can create safety incidents, which does not satisfy the requirement to map external errors into bounded user-visible outcomes. A suitable test of the correct implementation is: Test forbidden and not-found responses and inspect user-facing status and traces. This catches the particular incident where the agent tells a technician a repair was filed although the API returned 403.

C:

Users would believe unsupported operations completed. This shortcut can leave the following observable failure unresolved: The agent tells a technician a repair was filed although the API returned 403. A proper deployment instead must map external errors into bounded user-visible outcomes; the missing capability would leave the deployment exposed to the agent tells a technician a repair was filed although the API returned 403. The right approach is to handle status codes distinctly, preserve correlation IDs, and never invent success after failure, rather than relying on this unrelated shortcut. The distinguishing test is: Test forbidden and not-found responses and inspect user-facing status and traces.

D:

Identity changes are not a legitimate error recovery mechanism, so it cannot establish map external errors into bounded user-visible outcomes. The working design instead needs to handle status codes distinctly, preserve correlation IDs, and never invent success after failure. A strong boundary test is: Recover from a transient 503 without duplicating maintenance tickets. This matters because different failures have different safe retry rules and operational consequences.

Question 184

During incident response, telemetry shows: The agent tells a technician a repair was filed although the API returned 403. The feature exists to map external errors into bounded user-visible outcomes. What is the most relevant diagnosis?

  1. Tool-error mapping incorrectly equated failure with success.
  2. The orchestrator failed to preserve source provenance and role-specific outputs.
  3. Memory lookup lacks effective principal or tenant filtering.
  4. The agent retried a timed-out write without idempotency or reconciliation.

Explanation

Correct Answer

A:

The failure that the agent tells a technician a repair was filed although the API returned 403 points to tool-error mapping incorrectly equated failure with success. The direct functional check is: Test forbidden and not-found responses and inspect user-facing status and traces. Then test the independent boundary—recover from a transient 503 without duplicating maintenance tickets—to distinguish a genuine repair from a passing routine response.

Incorrect Answers

B:

If the orchestrator failed to preserve source provenance and role-specific outputs, then the expected symptom is the final answer cites a translation agent for a claim that required retrieval; this incident instead reports that the agent tells a technician a repair was filed although the API returned 403. The nearer diagnosis is tool-error mapping incorrectly equated failure with success; to distinguish it from the alternative of the orchestrator failed to preserve source provenance and role-specific outputs, test the specific service behavior: Test forbidden and not-found responses and inspect user-facing status and traces.

C:

If memory lookup lacks effective principal or tenant filtering, then the expected symptom is user A receives a preference only user B saved; this incident instead reports that the agent tells a technician a repair was filed although the API returned 403. The nearer diagnosis is tool-error mapping incorrectly equated failure with success; to distinguish it from the alternative of memory lookup lacks effective principal or tenant filtering, test the specific service behavior: Test forbidden and not-found responses and inspect user-facing status and traces.

D:

If the agent retried a timed-out write without idempotency or reconciliation, then the expected symptom is two reservations appear for the same trip after one user request; this incident instead reports that the agent tells a technician a repair was filed although the API returned 403. The nearer diagnosis is tool-error mapping incorrectly equated failure with success; to distinguish it from the alternative of the agent retried a timed-out write without idempotency or reconciliation, test the specific service behavior: Test forbidden and not-found responses and inspect user-facing status and traces.

Question 185

The project team needs production acceptance evidence. Its application has this need: A field-service agent must call a maintenance API that returns tool-specific errors such as missing asset or forbidden location. The required outcome is to map external errors into bounded user-visible outcomes. Select TWO checks that test this outcome.

  1. Return a generic success message after a tool returns 403 or 404 and log only the transport status.
  2. Retry an operation on every 503 without a bounded retry policy or idempotency check.
  3. Test only a normal 200 response and infer that failure mapping and 503 recovery are correct.
  4. Test forbidden and not-found responses and inspect user-facing status and traces.
  5. Recover from a transient 503 without duplicating maintenance tickets.

Explanation

Correct Answers

D:

Test forbidden and not-found responses and inspect user-facing status and traces measures the actual outcome needed to map external errors into bounded user-visible outcomes. It would expose the specific failure where the agent tells a technician a repair was filed although the API returned 403. Pair it with recover from a transient 503 without duplicating maintenance tickets to check both successful behavior and the remaining operational boundary.

E:

Recover from a transient 503 without duplicating maintenance tickets exercises a relevant negative or edge-case condition. It matters here because different failures have different safe retry rules and operational consequences. The complementary positive check—test forbidden and not-found responses and inspect user-facing status and traces—then confirms the intended business capability still works; neither observation alone settles the whole requirement.

Incorrect Answers

A:

A forbidden or absent resource is not a successful task completion. For this workload, the decisive functional evidence is: Test forbidden and not-found responses and inspect user-facing status and traces. The shortcut in this option could pass even when the agent tells a technician a repair was filed although the API returned 403.

B:

Uncontrolled retries can duplicate tickets or amplify outages. The deployment needs to map external errors into bounded user-visible outcomes. An independent test is: Recover from a transient 503 without duplicating maintenance tickets.

C:

Happy-path execution does not validate error semantics or safe retry behavior. The real incident to guard against is that the agent tells a technician a repair was filed although the API returned 403. This does not replace the case-specific validation: Recover from a transient 503 without duplicating maintenance tickets.

Question 186

The team has implemented the selected Azure design for this scenario: A field-service agent must call a maintenance API that returns tool-specific errors such as missing asset or forbidden location. It intends to map external errors into bounded user-visible outcomes. Which remaining operational limitation is specifically relevant?

  1. Different failures have different safe retry rules and operational consequences.
  2. Transcription and diarization errors can propagate even when the model cites real timestamps.
  3. Typed parsing prevents some errors but cannot verify a compromised upstream pricing service.
  4. Even isolated retrieval can expose data through shared logs or cached responses.

Explanation

Correct Answer

A:

Even after the team chooses to handle status codes distinctly, preserve correlation IDs, and never invent success after failure, different failures have different safe retry rules and operational consequences. This matters for a field-service agent must call a maintenance API that returns tool-specific errors such as missing asset or forbidden location. The edge-case check is: Recover from a transient 503 without duplicating maintenance tickets. The regular success check alone—test forbidden and not-found responses and inspect user-facing status and traces—cannot establish that this risk is controlled.

Incorrect Answers

B:

Transcription and diarization errors can propagate even when the model cites real timestamps could be important if a meeting-summary tool must cite exact supporting transcript segments and avoid filling in inaudible passages, but the actual use case is a field-service agent must call a maintenance API that returns tool-specific errors such as missing asset or forbidden location. The closest operational risk is different failures have different safe retry rules and operational consequences; evaluate that instead of transcription and diarization errors can propagate even when the model cites real timestamps by testing: Recover from a transient 503 without duplicating maintenance tickets.

C:

Typed parsing prevents some errors but cannot verify a compromised upstream pricing service could be important if a product assistant receives free-text tool responses but needs to pass prices as validated numbers, but the actual use case is a field-service agent must call a maintenance API that returns tool-specific errors such as missing asset or forbidden location. The closest operational risk is different failures have different safe retry rules and operational consequences; evaluate that instead of typed parsing prevents some errors but cannot verify a compromised upstream pricing service by testing: Recover from a transient 503 without duplicating maintenance tickets.

D:

Even isolated retrieval can expose data through shared logs or cached responses could be important if a legal retrieval assistant must respond to a question about a contract without using evidence from other clients, but the actual use case is a field-service agent must call a maintenance API that returns tool-specific errors such as missing asset or forbidden location. The closest operational risk is different failures have different safe retry rules and operational consequences; evaluate that instead of even isolated retrieval can expose data through shared logs or cached responses by testing: Recover from a transient 503 without duplicating maintenance tickets.

Question 187

An engineering review identifies this need: An agent invokes a remote MCP tool that exposes internal document search. To authenticate tool connections and scope allowed operations, which configuration should the team choose?

  1. Expose an unauthenticated public MCP endpoint for sensitive search.
  2. Allow an LLM to discover arbitrary MCP server URLs at runtime.
  3. Grant all MCP tools access to production storage-admin secrets.
  4. Use supported managed identity or OAuth permissions with an allowlisted tool catalog.

Explanation

Correct Answer

D:

MCP connectivity should carry controlled identity and explicit tool authorization; for this workload the decision is to use supported managed identity or OAuth permissions with an allowlisted tool catalog. The acceptance result is: Call an allowed search operation using the agent's configured identity. An independent edge-case check is: Attempt a disallowed write tool and verify it cannot be discovered or invoked. That distinction matters because tool whitelists must be revised as integrations add methods or permissions.

Incorrect Answers

A:

Public access bypasses document permissions, which does not satisfy the requirement to authenticate tool connections and scope allowed operations. A suitable test of the correct implementation is: Call an allowed search operation using the agent's configured identity. This catches the particular incident where the agent connects but search fails with authorization denied for a permitted project.

B:

Untrusted destinations expand attack surface. This shortcut can leave the following observable failure unresolved: The agent connects but search fails with authorization denied for a permitted project. A proper deployment instead must authenticate tool connections and scope allowed operations; the missing capability would leave the deployment exposed to the agent connects but search fails with authorization denied for a permitted project. The right approach is to use supported managed identity or OAuth permissions with an allowlisted tool catalog, rather than relying on this unrelated shortcut. The distinguishing test is: Call an allowed search operation using the agent's configured identity.

C:

Broad secrets undermine tool isolation, so it cannot establish authenticate tool connections and scope allowed operations. The working design instead needs to use supported managed identity or OAuth permissions with an allowlisted tool catalog. A strong boundary test is: Attempt a disallowed write tool and verify it cannot be discovered or invoked. This matters because tool whitelists must be revised as integrations add methods or permissions.

Question 188

After rollout, the team observes: The agent connects but search fails with authorization denied for a permitted project. The intended design was to authenticate tool connections and scope allowed operations. Which root cause should be investigated first?

  1. The pre-execution authorization gate counted missing or duplicate approvals as valid.
  2. The orchestrator failed to preserve source provenance and role-specific outputs.
  3. The MCP identity or delegated permissions lack the required target-resource role.
  4. Memory lookup lacks effective principal or tenant filtering.

Explanation

Correct Answer

C:

The failure that the agent connects but search fails with authorization denied for a permitted project points to the MCP identity or delegated permissions lack the required target-resource role. The direct functional check is: Call an allowed search operation using the agent's configured identity. Then test the independent boundary—attempt a disallowed write tool and verify it cannot be discovered or invoked—to distinguish a genuine repair from a passing routine response.

Incorrect Answers

A:

If the pre-execution authorization gate counted missing or duplicate approvals as valid, then the expected symptom is a payment executes with one approval despite a two-person rule; this incident instead reports that the agent connects but search fails with authorization denied for a permitted project. The nearer diagnosis is the MCP identity or delegated permissions lack the required target-resource role; to distinguish it from the alternative of the pre-execution authorization gate counted missing or duplicate approvals as valid, test the specific service behavior: Call an allowed search operation using the agent's configured identity.

B:

If the orchestrator failed to preserve source provenance and role-specific outputs, then the expected symptom is the final answer cites a translation agent for a claim that required retrieval; this incident instead reports that the agent connects but search fails with authorization denied for a permitted project. The nearer diagnosis is the MCP identity or delegated permissions lack the required target-resource role; to distinguish it from the alternative of the orchestrator failed to preserve source provenance and role-specific outputs, test the specific service behavior: Call an allowed search operation using the agent's configured identity.

D:

If memory lookup lacks effective principal or tenant filtering, then the expected symptom is user A receives a preference only user B saved; this incident instead reports that the agent connects but search fails with authorization denied for a permitted project. The nearer diagnosis is the MCP identity or delegated permissions lack the required target-resource role; to distinguish it from the alternative of memory lookup lacks effective principal or tenant filtering, test the specific service behavior: Call an allowed search operation using the agent's configured identity.

Question 189

A release reviewer considers this scenario: An agent invokes a remote MCP tool that exposes internal document search. The objective is to authenticate tool connections and scope allowed operations. Which TWO validation outcomes directly establish that behavior? Select TWO answers.

  1. Hide a write tool from the UI but leave it callable through a model-selected tool name.
  2. Confirm an allowed search works but never attempt an off-allowlist write tool.
  3. Attempt a disallowed write tool and verify it cannot be discovered or invoked.
  4. Call an allowed search operation using the agent's configured identity.
  5. Verify that the agent promises not to call write tools and trust that text instead of enforcement logs.

Explanation

Correct Answers

C:

Attempt a disallowed write tool and verify it cannot be discovered or invoked exercises a relevant negative or edge-case condition. It matters here because tool whitelists must be revised as integrations add methods or permissions. The complementary positive check—call an allowed search operation using the agent's configured identity—then confirms the intended business capability still works; neither observation alone settles the whole requirement.

D:

Call an allowed search operation using the agent's configured identity measures the actual outcome needed to authenticate tool connections and scope allowed operations. It would expose the specific failure where the agent connects but search fails with authorization denied for a permitted project. Pair it with attempt a disallowed write tool and verify it cannot be discovered or invoked to check both successful behavior and the remaining operational boundary.

Incorrect Answers

A:

Presentation restrictions do not enforce a server-side allowlist. For this workload, the decisive functional evidence is: Call an allowed search operation using the agent's configured identity. The shortcut in this option could pass even when the agent connects but search fails with authorization denied for a permitted project.

B:

An allowed-call test alone cannot demonstrate denied tool discovery or invocation. The deployment needs to authenticate tool connections and scope allowed operations. An independent test is: Attempt a disallowed write tool and verify it cannot be discovered or invoked.

E:

Instructions cannot substitute for actual tool permission enforcement. The real incident to guard against is that the agent connects but search fails with authorization denied for a permitted project. This does not replace the case-specific validation: Attempt a disallowed write tool and verify it cannot be discovered or invoked.

Question 190

A team implementing Microsoft Foundry faces this scenario: A multi-agent workflow has a planner and execution agent with conflicting proposed actions. The nonnegotiable requirement is to reconcile intermediate decisions before irreversible execution. Which design choice best meets it?

  1. Require the planner to produce a structured plan and the executor to verify constraints before tool calls.
  2. Discard all intermediate state before executing multi-step actions.
  3. Allow both agents to execute independent writes on the same records.
  4. Treat a persuasive planner message as proof the executor has permissions.

Explanation

Correct Answer

A:

A plan is advisory until the execution boundary validates it; for this workload the decision is to require the planner to produce a structured plan and the executor to verify constraints before tool calls. The acceptance result is: Submit a plan with a forbidden delete and verify no write occurs. An independent edge-case check is: Run a permitted plan and compare actual tool parameters to approved plan fields. That distinction matters because planner and executor can share incorrect assumptions even when their handoff is structured.

Incorrect Answers

B:

Without context, the executor cannot verify plan parameters, which does not satisfy the requirement to reconcile intermediate decisions before irreversible execution. A suitable test of the correct implementation is: Submit a plan with a forbidden delete and verify no write occurs. This catches the particular incident where an executor deletes an asset although the planner listed it as preserve-only.

C:

Concurrent uncoordinated actions can conflict. This shortcut can leave the following observable failure unresolved: An executor deletes an asset although the planner listed it as preserve-only. A proper deployment instead must reconcile intermediate decisions before irreversible execution; the missing capability would leave the deployment exposed to an executor deletes an asset although the planner listed it as preserve-only. The right approach is to require the planner to produce a structured plan and the executor to verify constraints before tool calls, rather than relying on this unrelated shortcut. The distinguishing test is: Submit a plan with a forbidden delete and verify no write occurs.

D:

Text does not grant tool authorization, so it cannot establish reconcile intermediate decisions before irreversible execution. The working design instead needs to require the planner to produce a structured plan and the executor to verify constraints before tool calls. A strong boundary test is: Run a permitted plan and compare actual tool parameters to approved plan fields. This matters because planner and executor can share incorrect assumptions even when their handoff is structured.

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

SPECIAL OFFER: GET 10% OFF
This is ONE TIME OFFER

You save
10%

Enter Your Email Address to Receive Your 10% Off Discount Code

SPECIAL OFFER: GET 10% OFF

You save
10%

Use Discount Code:

A confirmation link was sent to your e-mail.

Please check your mailbox for a message from support@examlabs.com and follow the directions.

Download Free Demo of VCE Exam Simulator

Experience Avanset VCE Exam Simulator for yourself.

Simply submit your email address below to get started with our interactive software demo of your free trial.

  • Realistic exam simulation and exam editor with preview functions
  • Whole exam in a single file with several different question types
  • Customizable exam-taking mode & detailed score reports