Microsoft AI-103 (Developing AI Apps and Agents on Azure) Practice Test 14

Question 130

A development group must stop an agent's broad language permissions from turning into unrestricted business-system privileges. What is the most appropriate Azure AI implementation?

  1. Enforce per-tool authorization independently
  2. Separate review and execution responsibilities
  3. Use an adversarial safety test suite before release
  4. Review false positives after moderation updates

Explanation

Correct Answer

A:

For the stated requirement to stop an agent's broad language permissions from turning into unrestricted business-system privileges, tools must authorize each action using the caller's allowed scope, regardless of what the LLM requests; natural-language rules are insufficient; this is the mechanism that the system must implement.

Incorrect Answers

B:

For a system whose objective is to stop an agent from approving its own high-risk recommendation, distinct approval identity or review service introduces independent oversight; a second prompt in the same privileged agent is not separation of duties; this does not directly meet the required outcome to stop an agent's broad language permissions from turning into unrestricted business-system privileges.

C:

For a system whose objective is to discover harmful outputs and instruction-bypass failures across diverse user and tool inputs, red-team cases exercise guardrail boundaries under realistic attacks, unlike happy-path functional tests alone; this does not directly meet the required outcome to stop an agent's broad language permissions from turning into unrestricted business-system privileges.

D:

For a system whose objective is to avoid treating all newly flagged legitimate customer requests as genuine safety threats, human-reviewed labeled examples distinguish benign business terms from actual prohibited content; thresholds and policy need measured tuning; this does not directly meet the required outcome to stop an agent's broad language permissions from turning into unrestricted business-system privileges.

Question 131

Production telemetry reveals this failure: Auditors cannot reproduce why an agent gave a policy answer after the underlying files were updated. Which change most directly corrects it?

  1. Enable appropriate indirect-injection screening and isolate retrieved text as untrusted data
  2. Audit disputed blocks, label representative examples, and tune supported thresholds carefully
  3. Add attack-oriented scenarios for jailbreaks, data leakage, and unsafe tool use before release
  4. Persist source-version identifiers and retrieval metadata with each audited decision

Explanation

Correct Answer

D:

The fault signal is that auditors cannot reproduce why an agent gave a policy answer after the underlying files were updated; the targeted fix works because document IDs, timestamps, and citation spans let reviewers reconstruct the evidence used at response time; later source edits should not erase the audit trail, restoring the ability to show exactly which source version supported a regulated answer.

Incorrect Answers

A:

If the failing function were to prevent an untrusted document from instructing an agent to ignore its governing rules, this change could be relevant because prompt Shields and related controls can detect indirect instructions embedded in lower-trust content; document text remains evidence, not authority; the actual symptom is that auditors cannot reproduce why an agent gave a policy answer after the underlying files were updated, which calls for a different repair.

B:

If the failing function were to avoid treating all newly flagged legitimate customer requests as genuine safety threats, this change could be relevant because human-reviewed labeled examples distinguish benign business terms from actual prohibited content; thresholds and policy need measured tuning; the actual symptom is that auditors cannot reproduce why an agent gave a policy answer after the underlying files were updated, which calls for a different repair.

C:

If the failing function were to discover harmful outputs and instruction-bypass failures across diverse user and tool inputs, this change could be relevant because red-team cases exercise guardrail boundaries under realistic attacks, unlike happy-path functional tests alone; the actual symptom is that auditors cannot reproduce why an agent gave a policy answer after the underlying files were updated, which calls for a different repair.

Question 132

An operator can reproduce this error: A refund agent can issue payouts without a human review even when evidence is contradictory. Which targeted repair is justified?

  1. Filter retrieval to documents authorized for the current caller and enforce access at the source
  2. Put a checked approval state between proposal and execution for payment-changing tools
  3. Redact sensitive fields at collection and restrict trace access according to retention policy
  4. Enforce authorization at tool execution and remove privileges not required by the agent role
  5. Configure and test the appropriate input and output harm guardrails for the deployed model

Explanation

Correct Answer

B:

The fault signal is that a refund agent can issue payouts without a human review even when evidence is contradictory; the targeted fix works because human-in-the-loop approval gates restrict irreversible operations even if the agent's reasoning or external evidence is wrong, restoring the ability to keep payment or account-changing tool calls under an explicit human authorization step.

Incorrect Answers

A:

If the failing function were to prevent an agent from citing restricted HR content to ordinary staff, this change could be relevant because security trimming or permission-aware retrieval applies user access rules before private passages reach generation; the actual symptom is that a refund agent can issue payouts without a human review even when evidence is contradictory, which calls for a different repair.

C:

If the failing function were to keep traces useful without retaining customer identifiers or secrets in operational logs, this change could be relevant because PII-aware redaction and minimal logging reduce privacy risk; collecting full prompts in diagnostic systems is not inherently necessary; the actual symptom is that a refund agent can issue payouts without a human review even when evidence is contradictory, which calls for a different repair.

D:

If the failing function were to stop an agent's broad language permissions from turning into unrestricted business-system privileges, this change could be relevant because tools must authorize each action using the caller's allowed scope, regardless of what the LLM requests; natural-language rules are insufficient; the actual symptom is that a refund agent can issue payouts without a human review even when evidence is contradictory, which calls for a different repair.

E:

If the failing function were to limit harmful user prompts and unsafe assistant responses according to an approved safety policy, this change could be relevant because content filters classify supported harm categories at the inference boundary; thresholds require evaluation for false positives and misses; the actual symptom is that a refund agent can issue payouts without a human review even when evidence is contradictory, which calls for a different repair.

Question 133

During a regression test, the following unexpected result occurs: After a filter update, routine medical-benefits enquiries are blocked disproportionately. What should the engineer do?

  1. Add attack-oriented scenarios for jailbreaks, data leakage, and unsafe tool use before release
  2. Audit disputed blocks, label representative examples, and tune supported thresholds carefully
  3. Enforce authorization at tool execution and remove privileges not required by the agent role
  4. Configure and test the appropriate input and output harm guardrails for the deployed model

Explanation

Correct Answer

B:

The fault signal is that after a filter update, routine medical-benefits enquiries are blocked disproportionately; the targeted fix works because human-reviewed labeled examples distinguish benign business terms from actual prohibited content; thresholds and policy need measured tuning, restoring the ability to avoid treating all newly flagged legitimate customer requests as genuine safety threats.

Incorrect Answers

A:

If the failing function were to discover harmful outputs and instruction-bypass failures across diverse user and tool inputs, this change could be relevant because red-team cases exercise guardrail boundaries under realistic attacks, unlike happy-path functional tests alone; the actual symptom is that after a filter update, routine medical-benefits enquiries are blocked disproportionately, which calls for a different repair.

C:

If the failing function were to stop an agent's broad language permissions from turning into unrestricted business-system privileges, this change could be relevant because tools must authorize each action using the caller's allowed scope, regardless of what the LLM requests; natural-language rules are insufficient; the actual symptom is that after a filter update, routine medical-benefits enquiries are blocked disproportionately, which calls for a different repair.

D:

If the failing function were to limit harmful user prompts and unsafe assistant responses according to an approved safety policy, this change could be relevant because content filters classify supported harm categories at the inference boundary; thresholds require evaluation for false positives and misses; the actual symptom is that after a filter update, routine medical-benefits enquiries are blocked disproportionately, which calls for a different repair.

Question 134

During a regression test, the following unexpected result occurs: An answer source contains 'ignore previous instructions' and the agent starts following that embedded command. What should the engineer do?

  1. Filter retrieval to documents authorized for the current caller and enforce access at the source
  2. Configure and test the appropriate input and output harm guardrails for the deployed model
  3. Enable appropriate indirect-injection screening and isolate retrieved text as untrusted data
  4. Redact sensitive fields at collection and restrict trace access according to retention policy

Explanation

Correct Answer

C:

The fault signal is that an answer source contains 'ignore previous instructions' and the agent starts following that embedded command; the targeted fix works because prompt Shields and related controls can detect indirect instructions embedded in lower-trust content; document text remains evidence, not authority, restoring the ability to prevent an untrusted document from instructing an agent to ignore its governing rules.

Incorrect Answers

A:

If the failing function were to prevent an agent from citing restricted HR content to ordinary staff, this change could be relevant because security trimming or permission-aware retrieval applies user access rules before private passages reach generation; the actual symptom is that an answer source contains 'ignore previous instructions' and the agent starts following that embedded command, which calls for a different repair.

B:

If the failing function were to limit harmful user prompts and unsafe assistant responses according to an approved safety policy, this change could be relevant because content filters classify supported harm categories at the inference boundary; thresholds require evaluation for false positives and misses; the actual symptom is that an answer source contains 'ignore previous instructions' and the agent starts following that embedded command, which calls for a different repair.

D:

If the failing function were to keep traces useful without retaining customer identifiers or secrets in operational logs, this change could be relevant because PII-aware redaction and minimal logging reduce privacy risk; collecting full prompts in diagnostic systems is not inherently necessary; the actual symptom is that an answer source contains 'ignore previous instructions' and the agent starts following that embedded command, which calls for a different repair.

Question 135

Before approving the proposed action to remove or mask sensitive personal data before logging, which measurable acceptance check should be required?

  1. Evaluate precision and recall for the affected safety categories after adjustment
  2. Run a policy-labeled evaluation set measuring block rates and legitimate-request false positives
  3. Sample logs to verify redaction coverage and authorized retention behavior
  4. Test that execution is blocked without approval and that every approval is attributable

Explanation

Correct Answer

C:

The accepted change is to remove or mask sensitive personal data before logging; this evaluation checks whether it can keep traces useful without retaining customer identifiers or secrets in operational logs, since PII-aware redaction and minimal logging reduce privacy risk; collecting full prompts in diagnostic systems is not inherently necessary.

Incorrect Answers

A:

This test concerns the ability to avoid treating all newly flagged legitimate customer requests as genuine safety threats, for which human-reviewed labeled examples distinguish benign business terms from actual prohibited content; thresholds and policy need measured tuning; it does not establish whether the change to remove or mask sensitive personal data before logging works for the current release.

B:

This test concerns the ability to limit harmful user prompts and unsafe assistant responses according to an approved safety policy, for which content filters classify supported harm categories at the inference boundary; thresholds require evaluation for false positives and misses; it does not establish whether the change to remove or mask sensitive personal data before logging works for the current release.

D:

This test concerns the ability to keep payment or account-changing tool calls under an explicit human authorization step, for which human-in-the-loop approval gates restrict irreversible operations even if the agent's reasoning or external evidence is wrong; it does not establish whether the change to remove or mask sensitive personal data before logging works for the current release.

Question 136

An Azure AI workload passes basic health probes but has this issue: A loan recommendation system marks its own proposal approved without any independent reviewer. Which fix targets it?

  1. Set a maximum step count, tool-call budget, and explicit abort or handoff rules
  2. Filter retrieval to documents authorized for the current caller and enforce access at the source
  3. Require approval by a separate authorized reviewer before committing regulated outcomes
  4. Add attack-oriented scenarios for jailbreaks, data leakage, and unsafe tool use before release

Explanation

Correct Answer

C:

The fault signal is that a loan recommendation system marks its own proposal approved without any independent reviewer; the targeted fix works because distinct approval identity or review service introduces independent oversight; a second prompt in the same privileged agent is not separation of duties, restoring the ability to stop an agent from approving its own high-risk recommendation.

Incorrect Answers

A:

If the failing function were to prevent infinite tool loops and uncontrolled consumption in long-running workflows, this change could be relevant because execution limits on tool calls, runtime and cost bound what autonomy can do, independent of prompt compliance; the actual symptom is that a loan recommendation system marks its own proposal approved without any independent reviewer, which calls for a different repair.

B:

If the failing function were to prevent an agent from citing restricted HR content to ordinary staff, this change could be relevant because security trimming or permission-aware retrieval applies user access rules before private passages reach generation; the actual symptom is that a loan recommendation system marks its own proposal approved without any independent reviewer, which calls for a different repair.

D:

If the failing function were to discover harmful outputs and instruction-bypass failures across diverse user and tool inputs, this change could be relevant because red-team cases exercise guardrail boundaries under realistic attacks, unlike happy-path functional tests alone; the actual symptom is that a loan recommendation system marks its own proposal approved without any independent reviewer, which calls for a different repair.

Question 137

The support ticket describes this behavior: The team cannot explain why similar inputs were accepted last month but blocked after a model release. Which engineering change should come first?

  1. Set a maximum step count, tool-call budget, and explicit abort or handoff rules
  2. Audit disputed blocks, label representative examples, and tune supported thresholds carefully
  3. Require approval by a separate authorized reviewer before committing regulated outcomes
  4. Version the safety configuration with release artifacts and preserve change history

Explanation

Correct Answer

D:

The fault signal is that the team cannot explain why similar inputs were accepted last month but blocked after a model release; the targeted fix works because versioned guardrail policies make thresholds, blocklists and exception decisions reproducible in evaluations and incidents, restoring the ability to explain changes in moderation behavior across releases and support rollback.

Incorrect Answers

A:

If the failing function were to prevent infinite tool loops and uncontrolled consumption in long-running workflows, this change could be relevant because execution limits on tool calls, runtime and cost bound what autonomy can do, independent of prompt compliance; the actual symptom is that the team cannot explain why similar inputs were accepted last month but blocked after a model release, which calls for a different repair.

B:

If the failing function were to avoid treating all newly flagged legitimate customer requests as genuine safety threats, this change could be relevant because human-reviewed labeled examples distinguish benign business terms from actual prohibited content; thresholds and policy need measured tuning; the actual symptom is that the team cannot explain why similar inputs were accepted last month but blocked after a model release, which calls for a different repair.

C:

If the failing function were to stop an agent from approving its own high-risk recommendation, this change could be relevant because distinct approval identity or review service introduces independent oversight; a second prompt in the same privileged agent is not separation of duties; the actual symptom is that the team cannot explain why similar inputs were accepted last month but blocked after a model release, which calls for a different repair.

Question 138

After implementing a workflow to limit knowledge access to the caller's entitlements, the team needs a robust release gate. Which check should it choose?

  1. Audit that proposer and approver identities are distinct on completed decisions
  2. Red-team malicious instructions inside retrieved pages and record detected unsafe tool intentions
  3. Induce a circular task and verify deterministic termination and budget reporting
  4. Test cross-role queries for unauthorized citations and verify least-privilege results

Explanation

Correct Answer

D:

The accepted change is to limit knowledge access to the caller's entitlements; this evaluation checks whether it can prevent an agent from citing restricted HR content to ordinary staff, since security trimming or permission-aware retrieval applies user access rules before private passages reach generation.

Incorrect Answers

A:

This test concerns the ability to stop an agent from approving its own high-risk recommendation, for which distinct approval identity or review service introduces independent oversight; a second prompt in the same privileged agent is not separation of duties; it does not establish whether the change to limit knowledge access to the caller's entitlements works for the current release.

B:

This test concerns the ability to prevent an untrusted document from instructing an agent to ignore its governing rules, for which prompt Shields and related controls can detect indirect instructions embedded in lower-trust content; document text remains evidence, not authority; it does not establish whether the change to limit knowledge access to the caller's entitlements works for the current release.

C:

This test concerns the ability to prevent infinite tool loops and uncontrolled consumption in long-running workflows, for which execution limits on tool calls, runtime and cost bound what autonomy can do, independent of prompt compliance; it does not establish whether the change to limit knowledge access to the caller's entitlements works for the current release.

Question 139

The planned deployment includes a mechanism to constrain an autonomous agent's steps and budget. Which outcome should the QA team measure?

  1. Test cross-role queries for unauthorized citations and verify least-privilege results
  2. Induce a circular task and verify deterministic termination and budget reporting
  3. Measure attack success rate and incident severity across protected versions
  4. Audit that proposer and approver identities are distinct on completed decisions

Explanation

Correct Answer

B:

The accepted change is to constrain an autonomous agent's steps and budget; this evaluation checks whether it can prevent infinite tool loops and uncontrolled consumption in long-running workflows, since execution limits on tool calls, runtime and cost bound what autonomy can do, independent of prompt compliance.

Incorrect Answers

A:

This test concerns the ability to prevent an agent from citing restricted HR content to ordinary staff, for which security trimming or permission-aware retrieval applies user access rules before private passages reach generation; it does not establish whether the change to constrain an autonomous agent's steps and budget works for the current release.

C:

This test concerns the ability to discover harmful outputs and instruction-bypass failures across diverse user and tool inputs, for which red-team cases exercise guardrail boundaries under realistic attacks, unlike happy-path functional tests alone; it does not establish whether the change to constrain an autonomous agent's steps and budget works for the current release.

D:

This test concerns the ability to stop an agent from approving its own high-risk recommendation, for which distinct approval identity or review service introduces independent oversight; a second prompt in the same privileged agent is not separation of duties; it does not establish whether the change to constrain an autonomous agent's steps and budget works for the current release.

Question 140

After implementing a workflow to apply input and output harm filters, the team needs a robust release gate. Which check should it choose?

  1. Sample logs to verify redaction coverage and authorized retention behavior
  2. Induce a circular task and verify deterministic termination and budget reporting
  3. Run a policy-labeled evaluation set measuring block rates and legitimate-request false positives
  4. Red-team malicious instructions inside retrieved pages and record detected unsafe tool intentions

Explanation

Correct Answer

C:

The accepted change is to apply input and output harm filters; this evaluation checks whether it can limit harmful user prompts and unsafe assistant responses according to an approved safety policy, since content filters classify supported harm categories at the inference boundary; thresholds require evaluation for false positives and misses.

Incorrect Answers

A:

This test concerns the ability to keep traces useful without retaining customer identifiers or secrets in operational logs, for which PII-aware redaction and minimal logging reduce privacy risk; collecting full prompts in diagnostic systems is not inherently necessary; it does not establish whether the change to apply input and output harm filters works for the current release.

B:

This test concerns the ability to prevent infinite tool loops and uncontrolled consumption in long-running workflows, for which execution limits on tool calls, runtime and cost bound what autonomy can do, independent of prompt compliance; it does not establish whether the change to apply input and output harm filters works for the current release.

D:

This test concerns the ability to prevent an untrusted document from instructing an agent to ignore its governing rules, for which prompt Shields and related controls can detect indirect instructions embedded in lower-trust content; document text remains evidence, not authority; it does not establish whether the change to apply input and output harm filters works for the current release.

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