Microsoft AI-103 Practice Test 40 - Prepaway
Practice Exams:

Pass Microsoft AI-103 Exam in First Attempt Guaranteed!

Get 100% Latest Exam Questions, Accurate & Verified Answers!

30 Days Free Updates, Instant Download!

Verified By Experts
Premium File
$89.99 $98.99

AI-103 Premium ETE File


  • Premium File 187 Questions & Answers. Last Update: Nov 15, 2026
Whats Included:
  • Latest Questions
  • 100% Accurate Answers
  • Fast Exam Updates

Microsoft AI-103 Practice Test 40

Question 390

Production requirement: A financial RAG tool retrieves an external PDF containing text impersonating a new administrator message. Which TWO decisions would undermine the required system boundary? Select TWO answers.

  1. Separate retrieved untrusted content from system instructions, enforce tool and data policies outside the model, and test adversarial retrieval cases
  2. Include production secret values in the retrieved document context to make the model more helpful
  3. Let a retrieved article override developer tool-access rules when it claims to be a system instruction
  4. Known malicious retrieved snippets fail to redirect tool privileges or leak secrets while legitimate citation-based answers still succeed
  5. Inspect records that explain the affected process, then check the result described in this case: A financial RAG tool retrieves an external PDF containing text impersonating a new administrator message

Explanation

Correct Answer

B:

“Include production secret values in the retrieved document context to make the model more helpful” would violate the requirement.

C:

“Let a retrieved article override developer tool-access rules when it claims to be a system instruction” would violate the requirement.

Incorrect Answers

A:

“Separate retrieved untrusted content from system instructions, enforce tool and data policies outside the model, and test adversarial retrieval cases” is the controlled implementation, not an unsafe workaround. Confirm that known malicious retrieved snippets fail to redirect tool privileges or leak secrets while legitimate citation-based answers still succeed.

D:

“Known malicious retrieved snippets fail to redirect tool privileges or leak secrets while legitimate citation-based answers still succeed” is the required outcome of a working control, not an action that weakens it. Confirm that known malicious retrieved snippets fail to redirect tool privileges or leak secrets while legitimate citation-based answers still succeed.

E:

“Inspect records that explain the affected process, then check the result described in this case: A financial RAG tool retrieves an external PDF containing text impersonating a new administrator message” adds a relevant safeguard or verification step rather than weakening the solution.

Question 391

A team is investigating this task: A travel booking agent receives an airport webpage that tries to redirect its payment tool to an unrelated account. Which design decision is justified by the failure and the required outcome?

  1. Apply the proposed setting to all connected workloads before testing the case-specific boundary: a travel booking agent receives an airport webpage that tries to redirect its payment tool to an unrelated account
  2. Separate retrieved untrusted content from system instructions, enforce tool and data policies outside the model, and test adversarial retrieval cases
  3. Check only clean documents rather than adversarial input cases
  4. Promote every cited passage to a trusted instruction without verifying its source

Explanation

Correct Answer

B:

A travel booking agent receives an airport webpage that tries to redirect its payment tool to an unrelated account. Retrieved documents and tool outputs are data, not higher-priority instructions; application-side authorization must survive malicious text in the retrieved context. Known malicious retrieved snippets fail to redirect tool privileges or leak secrets while legitimate citation-based answers still succeed.

Incorrect Answers

A:

“Apply the proposed setting to all connected workloads before testing the case-specific boundary: a travel booking agent receives an airport webpage that tries to redirect its payment tool to an unrelated” does not implement the control that the facts call for. Retrieved documents and tool outputs are data, not higher-priority instructions; application-side authorization must survive malicious text in the retrieved context. Instead, separate retrieved untrusted content from system instructions, enforce tool and data policies outside the model, and test adversarial retrieval cases.

C:

A travel booking agent receives an airport webpage that tries to redirect its payment tool to an unrelated account. “Check only clean documents rather than adversarial input cases” does not implement the control that the facts call for. Retrieved documents and tool outputs are data, not higher-priority instructions; application-side authorization must survive malicious text in the retrieved context. Instead, separate retrieved untrusted content from system instructions, enforce tool and data policies outside the model, and test adversarial retrieval cases.

D:

Because a team is investigating this task: A travel booking agent receives an airport webpage that tries to redirect its payment tool to an unrelated account, “Promote every cited passage to a trusted instruction without verifying its source” does not implement the control that the facts call for. Retrieved documents and tool outputs are data, not higher-priority instructions; application-side authorization must survive malicious text in the retrieved context. Instead, separate retrieved untrusted content from system instructions, enforce tool and data policies outside the model, and test adversarial retrieval cases.

Question 392

Situation: A travel booking agent receives an airport webpage that tries to redirect its payment tool to an unrelated account. Which underlying technical principle explains the recommended implementation?

  1. The fact that the team started work is enough to prove correctness for this exact case: a travel booking agent receives an airport webpage that tries to redirect its payment tool to an unrelated account
  2. Successful provisioning by itself proves all security, output, and operational constraints are satisfied
  3. The most powerful or least restricted model is always the safest choice regardless of task conditions
  4. Retrieved documents and tool outputs are data, not higher-priority instructions; application-side authorization must survive malicious text in the retrieved context

Explanation

Correct Answer

D:

Retrieved documents and tool outputs are data, not higher-priority instructions; application-side authorization must survive malicious text in the retrieved context. Separate retrieved untrusted content from system instructions, enforce tool and data policies outside the model, and test adversarial retrieval cases. The decisive evidence is that known malicious retrieved snippets fail to redirect tool privileges or leak secrets while legitimate citation-based answers still succeed.

Incorrect Answers

A:

“The fact that the team started work is enough to prove correctness for this exact case: a travel booking agent receives an airport webpage that tries to redirect its payment tool to an unrelated account” does not fully explain the decisive technical boundary in this case. Retrieved documents and tool outputs are data, not higher-priority instructions; application-side authorization must survive malicious text in the retrieved context.

B:

Because a team is investigating this task: A travel booking agent receives an airport webpage that tries to redirect its payment tool to an unrelated account, “Successful provisioning by itself proves all security, output, and operational constraints are satisfied” does not fully explain the decisive technical boundary in this case. Retrieved documents and tool outputs are data, not higher-priority instructions; application-side authorization must survive malicious text in the retrieved context. The implementation must separate retrieved untrusted content from system instructions, enforce tool and data policies outside the model, and test adversarial retrieval cases.

C:

A team is investigating this task: A travel booking agent receives an airport webpage that tries to redirect its payment tool to an unrelated account. “The most powerful or least restricted model is always the safest choice regardless of task conditions” does not fully explain the decisive technical boundary in this case. Retrieved documents and tool outputs are data, not higher-priority instructions; application-side authorization must survive malicious text in the retrieved context. The implementation must separate retrieved untrusted content from system instructions, enforce tool and data policies outside the model, and test adversarial retrieval cases.

Question 393

Architecture case: A travel booking agent receives an airport webpage that tries to redirect its payment tool to an unrelated account. What observable acceptance condition should be required before closing the task?

  1. A project dashboard remains green, although the affected request, output and permissions were not checked
  2. Known malicious retrieved snippets fail to redirect tool privileges or leak secrets while legitimate citation-based answers still succeed
  3. The operator reports success without reproducing the observed behavior: a travel booking agent receives an airport webpage that tries to redirect its payment tool to an unrelated account
  4. The team closes a change ticket without replaying a representative positive and negative test

Explanation

Correct Answer

B:

A travel booking agent receives an airport webpage that tries to redirect its payment tool to an unrelated account. A travel booking agent receives an airport webpage that tries to redirect its payment tool to an unrelated account. Retrieved documents and tool outputs are data, not higher-priority instructions; application-side authorization must survive malicious text in the retrieved context. Known malicious retrieved snippets fail to redirect tool privileges or leak secrets while legitimate citation-based answers still succeed.

Incorrect Answers

A:

A team is investigating this task: A travel booking agent receives an airport webpage that tries to redirect its payment tool to an unrelated account. “A project dashboard remains green, although the affected request, output and permissions were not checked” would not meet the requirement. Known malicious retrieved snippets fail to redirect tool privileges or leak secrets while legitimate citation-based answers still succeed.

C:

“The operator reports success without reproducing the observed behavior: a travel booking agent receives an airport webpage that tries to redirect its payment tool to an unrelated account” alone does not establish all the required behavior. Known malicious retrieved snippets fail to redirect tool privileges or leak secrets while legitimate citation-based answers still succeed.

D:

Because a team is investigating this task: A travel booking agent receives an airport webpage that tries to redirect its payment tool to an unrelated account, “The team closes a change ticket without replaying a representative positive and negative test” alone does not establish all the required behavior. Known malicious retrieved snippets fail to redirect tool privileges or leak secrets while legitimate citation-based answers still succeed.

Question 394

Production requirement: A travel booking agent receives an airport webpage that tries to redirect its payment tool to an unrelated account. Which first investigation is more useful than an unrelated health indicator?

  1. Assume the general service is at fault without checking the affected operation in this case: a travel booking agent receives an airport webpage that tries to redirect its payment tool to an unrelated account
  2. Restart every affected component before collecting request or tool traces
  3. Read only the broad service-health summary and skip actual workload evidence
  4. Inspect retrieval payload provenance, policy decision traces, attempted tool calls, rejected arguments and red-team evaluation outcomes

Explanation

Correct Answer

D:

A team is investigating this task: A travel booking agent receives an airport webpage that tries to redirect its payment tool to an unrelated account. Inspect retrieval payload provenance, policy decision traces, attempted tool calls, rejected arguments and red-team evaluation outcomes. Retrieved documents and tool outputs are data, not higher-priority instructions; application-side authorization must survive malicious text in the retrieved context. Known malicious retrieved snippets fail to redirect tool privileges or leak secrets while legitimate citation-based answers still succeed.

Incorrect Answers

A:

“Assume the general service is at fault without checking the affected operation in this case: a travel booking agent receives an airport webpage that tries to redirect its payment tool to an unrelated” would not meet the requirement. Inspect retrieval payload provenance, policy decision traces, attempted tool calls, rejected arguments and red-team evaluation outcomes.

B:

“Restart every affected component before collecting request or tool traces” would not be the most direct first check for this failure. Inspect retrieval payload provenance, policy decision traces, attempted tool calls, rejected arguments and red-team evaluation outcomes. Retrieved documents and tool outputs are data, not higher-priority instructions; application-side authorization must survive malicious text in the retrieved context.

C:

Because a team is investigating this task: A travel booking agent receives an airport webpage that tries to redirect its payment tool to an unrelated account, “Read only the broad service-health summary and skip actual workload evidence” would not meet the requirement. Inspect retrieval payload provenance, policy decision traces, attempted tool calls, rejected arguments and red-team evaluation outcomes. Retrieved documents and tool outputs are data, not higher-priority instructions; application-side authorization must survive malicious text in the retrieved context.

Question 395

A team is investigating this task: A travel booking agent receives an airport webpage that tries to redirect its payment tool to an unrelated account. Which TWO actions could conceal a serious failure while appearing faster? Select TWO answers.

  1. Inspect records that explain the affected process, then check the result described in this case: A travel booking agent receives an airport webpage that tries to redirect its payment tool to an unrelated account
  2. Known malicious retrieved snippets fail to redirect tool privileges or leak secrets while legitimate citation-based answers still succeed
  3. Let a retrieved article override developer tool-access rules when it claims to be a system instruction
  4. Include production secret values in the retrieved document context to make the model more helpful
  5. Separate retrieved untrusted content from system instructions, enforce tool and data policies outside the model, and test adversarial retrieval cases

Explanation

Correct Answer

C:

A team is investigating this task: A travel booking agent receives an airport webpage that tries to redirect its payment tool to an unrelated account. “Let a retrieved article override developer tool-access rules when it claims to be a system instruction” would violate the requirement. Retrieved documents and tool outputs are data, not higher-priority instructions; application-side authorization must survive malicious text in the retrieved context.

D:

A team is investigating this task: A travel booking agent receives an airport webpage that tries to redirect its payment tool to an unrelated account. “Include production secret values in the retrieved document context to make the model more helpful” would violate the requirement. Retrieved documents and tool outputs are data, not higher-priority instructions; application-side authorization must survive malicious text in the retrieved context.

Incorrect Answers

A:

“Inspect records that explain the affected process, then check the result described in this case: A travel booking agent receives an airport webpage that tries to redirect its payment tool to an unrelated” adds a relevant safeguard or verification step rather than weakening the solution. Retrieved documents and tool outputs are data, not higher-priority instructions; application-side authorization must survive malicious text in the retrieved context.

B:

A team is investigating this task: A travel booking agent receives an airport webpage that tries to redirect its payment tool to an unrelated account. “Known malicious retrieved snippets fail to redirect tool privileges or leak secrets while legitimate citation-based answers still succeed” is the required outcome of a working control, not an action that weakens it. Retrieved documents and tool outputs are data, not higher-priority instructions; application-side authorization must survive malicious text in the retrieved context. Confirm that known malicious retrieved snippets fail to redirect tool privileges or leak secrets while legitimate citation-based answers still succeed.

E:

“Separate retrieved untrusted content from system instructions, enforce tool and data policies outside the model, and test adversarial retrieval cases” is the controlled implementation, not an unsafe workaround. Retrieved documents and tool outputs are data, not higher-priority instructions; application-side authorization must survive malicious text in the retrieved context.

Question 396

Situation: A billing agent must return invoice number, approved amount and an ISO date, and a wrong type caused a database write failure. Which implementation best meets the requirement while honoring the stated constraints?

  1. Assume the phrase output JSON ensures syntactic validity
  2. Disable required-field validation to improve completion rate
  3. Constrain the output to a versioned JSON schema and validate types and required fields before invoking downstream writes
  4. Apply the proposed setting to all connected workloads before testing the case-specific boundary: a billing agent must return invoice number, approved amount and an iso date, and a wrong type caused a database write failure

Explanation

Correct Answer

C:

An LLM producing plausible JSON text does not guarantee conforming types, required properties or safe downstream tool parameters. Confirm that every accepted completion parses against the schema with required properties while invalid records are retried or rejected without side effects. If the result fails, capture raw model response, parser errors, schema-version IDs, field types and rejected-tool-call counts.

Incorrect Answers

A:

Because a billing agent must return invoice number, approved amount and an ISO date, and a wrong type caused a database write failure, “Assume the phrase output JSON ensures syntactic validity” would not meet the requirement. An LLM producing plausible JSON text does not guarantee conforming types, required properties or safe downstream tool parameters. Instead, constrain the output to a versioned JSON schema and validate types and required fields before invoking downstream writes.

B:

A billing agent must return invoice number, approved amount and an ISO date, and a wrong type caused a database write failure. “Disable required-field validation to improve completion rate” would not meet the requirement. An LLM producing plausible JSON text does not guarantee conforming types, required properties or safe downstream tool parameters. Instead, constrain the output to a versioned JSON schema and validate types and required fields before invoking downstream writes.

D:

“Apply the proposed setting to all connected workloads before testing the case-specific boundary: a billing agent must return invoice number, approved amount and an iso date” does not implement the control that the facts call for. An LLM producing plausible JSON text does not guarantee conforming types, required properties or safe downstream tool parameters. Instead, constrain the output to a versioned JSON schema and validate types and required fields before invoking downstream writes.

Question 397

Architecture case: A billing agent must return invoice number, approved amount and an ISO date, and a wrong type caused a database write failure. Which design boundary is most relevant to this decision?

  1. The fact that the team started work is enough to prove correctness for this exact case: a billing agent must return invoice number, approved amount and an iso date, and a wrong type caused a database write failure
  2. Successful provisioning by itself proves all security, output, and operational constraints are satisfied
  3. The most powerful or least restricted model is always the safest choice regardless of task conditions
  4. An LLM producing plausible JSON text does not guarantee conforming types, required properties or safe downstream tool parameters

Explanation

Correct Answer

D:

Because a billing agent must return invoice number, approved amount and an ISO date, and a wrong type caused a database write failure, an LLM producing plausible JSON text does not guarantee conforming types, required properties or safe downstream tool parameters. Constrain the output to a versioned JSON schema and validate types and required fields before invoking downstream writes. The decisive evidence is that every accepted completion parses against the schema with required properties while invalid records are retried or rejected without side effects.

Incorrect Answers

A:

“The fact that the team started work is enough to prove correctness for this exact case: a billing agent must return invoice number, approved amount and an iso date” does not fully explain the decisive technical boundary in this case. An LLM producing plausible JSON text does not guarantee conforming types, required properties or safe downstream tool parameters. The implementation must constrain the output to a versioned JSON schema and validate types and required fields before invoking downstream writes.

B:

“Successful provisioning by itself proves all security, output, and operational constraints are satisfied” does not fully explain the decisive technical boundary in this case. An LLM producing plausible JSON text does not guarantee conforming types, required properties or safe downstream tool parameters. The implementation must constrain the output to a versioned JSON schema and validate types and required fields before invoking downstream writes.

C:

Because a billing agent must return invoice number, approved amount and an ISO date, and a wrong type caused a database write failure, “The most powerful or least restricted model is always the safest choice regardless of task conditions” does not fully explain the decisive technical boundary in this case. An LLM producing plausible JSON text does not guarantee conforming types, required properties or safe downstream tool parameters.

Question 398

Production requirement: A billing agent must return invoice number, approved amount and an ISO date, and a wrong type caused a database write failure. Which evidence is strongest for proving this requirement was met?

  1. The operator reports success without reproducing the observed behavior: a billing agent must return invoice number, approved amount and an iso date, and a wrong type caused a database write failure
  2. A project dashboard remains green, although the affected request, output and permissions were not checked
  3. Every accepted completion parses against the schema with required properties while invalid records are retried or rejected without side effects
  4. The team closes a change ticket without replaying a representative positive and negative test

Explanation

Correct Answer

C:

A billing agent must return invoice number, approved amount and an ISO date, and a wrong type caused a database write failure. An LLM producing plausible JSON text does not guarantee conforming types, required properties or safe downstream tool parameters. Every accepted completion parses against the schema with required properties while invalid records are retried or rejected without side effects.

Incorrect Answers

A:

“The operator reports success without reproducing the observed behavior: a billing agent must return invoice number, approved amount and an iso date, and a wrong type caused a database write failure” alone does not establish all the required behavior. Every accepted completion parses against the schema with required properties while invalid records are retried or rejected without side effects. An LLM producing plausible JSON text does not guarantee conforming types, required properties or safe downstream tool parameters.

B:

“A project dashboard remains green, although the affected request, output and permissions were not checked” would not meet the requirement. Every accepted completion parses against the schema with required properties while invalid records are retried or rejected without side effects.

D:

A billing agent must return invoice number, approved amount and an ISO date, and a wrong type caused a database write failure. “The team closes a change ticket without replaying a representative positive and negative test” alone does not establish all the required behavior. Every accepted completion parses against the schema with required properties while invalid records are retried or rejected without side effects.

Question 399

A team is investigating this task: A billing agent must return invoice number, approved amount and an ISO date, and a wrong type caused a database write failure. What targeted records should the responder inspect to distinguish configuration from runtime behavior?

  1. Restart every affected component before collecting request or tool traces
  2. Read only the broad service-health summary and skip actual workload evidence
  3. Capture raw model response, parser errors, schema-version IDs, field types and rejected-tool-call counts
  4. Assume the general service is at fault without checking the affected operation in this case: a billing agent must return invoice number, approved amount and an iso date, and a wrong type caused a database write failure

Explanation

Correct Answer

C:

Capture raw model response, parser errors, schema-version IDs, field types and rejected-tool-call counts. An LLM producing plausible JSON text does not guarantee conforming types, required properties or safe downstream tool parameters. Every accepted completion parses against the schema with required properties while invalid records are retried or rejected without side effects.

Incorrect Answers

A:

A billing agent must return invoice number, approved amount and an ISO date, and a wrong type caused a database write failure. “Restart every affected component before collecting request or tool traces” would not be the most direct first check for this failure. Capture raw model response, parser errors, schema-version IDs, field types and rejected-tool-call counts. An LLM producing plausible JSON text does not guarantee conforming types, required properties or safe downstream tool parameters.

B:

A billing agent must return invoice number, approved amount and an ISO date, and a wrong type caused a database write failure. “Read only the broad service-health summary and skip actual workload evidence” would not meet the requirement. Capture raw model response, parser errors, schema-version IDs, field types and rejected-tool-call counts.

D:

“Assume the general service is at fault without checking the affected operation in this case: a billing agent must return invoice number, approved amount and an iso date” would not meet the requirement. Capture raw model response, parser errors, schema-version IDs, field types and rejected-tool-call counts.

Question 400

Situation: A billing agent must return invoice number, approved amount and an ISO date, and a wrong type caused a database write failure. Which TWO proposed practices create material safety, correctness, or reliability risks? Select TWO answers.

  1. Silently coerce invalid amounts into arbitrary numeric defaults
  2. Inspect records that explain the affected process, then check the result described in this case: A billing agent must return invoice number, approved amount and an ISO date, and a wrong type caused a database write failure
  3. Pass free-form model text directly as trusted database update parameters
  4. Constrain the output to a versioned JSON schema and validate types and required fields before invoking downstream writes
  5. Every accepted completion parses against the schema with required properties while invalid records are retried or rejected without side effects

Explanation

Correct Answer

A:

“Silently coerce invalid amounts into arbitrary numeric defaults” would violate the requirement. An LLM producing plausible JSON text does not guarantee conforming types, required properties or safe downstream tool parameters.

C:

A billing agent must return invoice number, approved amount and an ISO date, and a wrong type caused a database write failure. “Pass free-form model text directly as trusted database update parameters” would violate the requirement. An LLM producing plausible JSON text does not guarantee conforming types, required properties or safe downstream tool parameters.

Incorrect Answers

B:

“Inspect records that explain the affected process, then check the result described in this case: A billing agent must return invoice number, approved amount and an ISO date” adds a relevant safeguard or verification step rather than weakening the solution. An LLM producing plausible JSON text does not guarantee conforming types, required properties or safe downstream tool parameters.

D:

“Constrain the output to a versioned JSON schema and validate types and required fields before invoking downstream writes” is the controlled implementation, not an unsafe workaround. An LLM producing plausible JSON text does not guarantee conforming types, required properties or safe downstream tool parameters.

E:

Because a billing agent must return invoice number, approved amount and an ISO date, and a wrong type caused a database write failure, “Every accepted completion parses against the schema with required properties while invalid records are retried or rejected without side effects” is the required outcome of a working control, not an action that weakens it. An LLM producing plausible JSON text does not guarantee conforming types, required properties or safe downstream tool parameters. Confirm that every accepted completion parses against the schema with required properties while invalid records are retried or rejected without side effects.

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