Microsoft AI-103 (Developing AI Apps and Agents on Azure) Practice Test 13

Question 120

A new customer-facing AI service must prevent an agent from citing restricted HR content to ordinary staff. Which approach should the team select first?

  1. Constrain an autonomous agent's steps and budget
  2. Limit knowledge access to the caller's entitlements
  3. Use prompt-injection detection on retrieved content
  4. Separate review and execution responsibilities

Explanation

Correct Answer

B:

For the stated requirement to prevent an agent from citing restricted HR content to ordinary staff, security trimming or permission-aware retrieval applies user access rules before private passages reach generation; this is the mechanism that the system must implement.

Incorrect Answers

A:

For a system whose objective is to prevent infinite tool loops and uncontrolled consumption in long-running workflows, execution limits on tool calls, runtime and cost bound what autonomy can do, independent of prompt compliance; this does not directly meet the required outcome to prevent an agent from citing restricted HR content to ordinary staff.

C:

For a system whose objective is to prevent an untrusted document from instructing an agent to ignore its governing rules, prompt Shields and related controls can detect indirect instructions embedded in lower-trust content; document text remains evidence, not authority; this does not directly meet the required outcome to prevent an agent from citing restricted HR content to ordinary staff.

D:

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 prevent an agent from citing restricted HR content to ordinary staff.

Question 121

A development group must prevent infinite tool loops and uncontrolled consumption in long-running workflows. What is the most appropriate Azure AI implementation?

  1. Require human approval for high-impact agent actions
  2. Constrain an autonomous agent's steps and budget
  3. Limit knowledge access to the caller's entitlements
  4. Review false positives after moderation updates
  5. Use documented safety-policy versioning

Explanation

Correct Answer

B:

For the stated requirement to prevent infinite tool loops and uncontrolled consumption in long-running workflows, execution limits on tool calls, runtime and cost bound what autonomy can do, independent of prompt compliance; this is the mechanism that the system must implement.

Incorrect Answers

A:

For a system whose objective is to keep payment or account-changing tool calls under an explicit human authorization step, human-in-the-loop approval gates restrict irreversible operations even if the agent's reasoning or external evidence is wrong; this does not directly meet the required outcome to prevent infinite tool loops and uncontrolled consumption in long-running workflows.

C:

For a system whose objective is to prevent an agent from citing restricted HR content to ordinary staff, security trimming or permission-aware retrieval applies user access rules before private passages reach generation; this does not directly meet the required outcome to prevent infinite tool loops and uncontrolled consumption in long-running workflows.

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 prevent infinite tool loops and uncontrolled consumption in long-running workflows.

E:

For a system whose objective is to explain changes in moderation behavior across releases and support rollback, versioned guardrail policies make thresholds, blocklists and exception decisions reproducible in evaluations and incidents; this does not directly meet the required outcome to prevent infinite tool loops and uncontrolled consumption in long-running workflows.

Question 122

A production AI solution is being scoped for production. The non-negotiable task is to stop an agent from approving its own high-risk recommendation. Which implementation fits?

  1. Separate review and execution responsibilities
  2. Use documented safety-policy versioning
  3. Limit knowledge access to the caller's entitlements
  4. Enforce per-tool authorization independently

Explanation

Correct Answer

A:

For the stated requirement 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 is the mechanism that the system must implement.

Incorrect Answers

B:

For a system whose objective is to explain changes in moderation behavior across releases and support rollback, versioned guardrail policies make thresholds, blocklists and exception decisions reproducible in evaluations and incidents; this does not directly meet the required outcome to stop an agent from approving its own high-risk recommendation.

C:

For a system whose objective is to prevent an agent from citing restricted HR content to ordinary staff, security trimming or permission-aware retrieval applies user access rules before private passages reach generation; this does not directly meet the required outcome to stop an agent from approving its own high-risk recommendation.

D:

For a system whose objective is 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 does not directly meet the required outcome to stop an agent from approving its own high-risk recommendation.

Question 123

The engineering change request is to use prompt-injection detection on retrieved content. What is the most direct acceptance criterion?

  1. Sample logs to verify redaction coverage and authorized retention behavior
  2. Red-team malicious instructions inside retrieved pages and record detected unsafe tool intentions
  3. Run a policy-labeled evaluation set measuring block rates and legitimate-request false positives
  4. Evaluate precision and recall for the affected safety categories after adjustment

Explanation

Correct Answer

B:

The accepted change is to use prompt-injection detection on retrieved content; this evaluation checks whether it can prevent an untrusted document from instructing an agent to ignore its governing rules, since prompt Shields and related controls can detect indirect instructions embedded in lower-trust content; document text remains evidence, not authority.

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 use prompt-injection detection on retrieved content works for the current release.

C:

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 use prompt-injection detection on retrieved content works for the current release.

D:

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 use prompt-injection detection on retrieved content works for the current release.

Question 124

The support ticket describes this behavior: A read-only assistant can invoke an account-deletion function because the tool endpoint trusts all agent calls. Which engineering change should come first?

  1. Enforce authorization at tool execution and remove privileges not required by the agent role
  2. Persist source-version identifiers and retrieval metadata with each audited decision
  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

A:

The fault signal is that a read-only assistant can invoke an account-deletion function because the tool endpoint trusts all agent calls; the targeted fix works because tools must authorize each action using the caller's allowed scope, regardless of what the LLM requests; natural-language rules are insufficient, restoring the ability to stop an agent's broad language permissions from turning into unrestricted business-system privileges.

Incorrect Answers

B:

If the failing function were to show exactly which source version supported a regulated answer, this change could be relevant 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; the actual symptom is that a read-only assistant can invoke an account-deletion function because the tool endpoint trusts all agent calls, 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 a read-only assistant can invoke an account-deletion function because the tool endpoint trusts all agent calls, 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 read-only assistant can invoke an account-deletion function because the tool endpoint trusts all agent calls, which calls for a different repair.

Question 125

The acceptance criteria for a production AI solution require it to avoid treating all newly flagged legitimate customer requests as genuine safety threats. Which design should be approved?

  1. Remove or mask sensitive personal data before logging
  2. Record provenance for retrieved evidence
  3. Review false positives after moderation updates
  4. Use an adversarial safety test suite before release

Explanation

Correct Answer

C:

For the stated requirement 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 is the mechanism that the system must implement.

Incorrect Answers

A:

For a system whose objective is to keep traces useful without retaining customer identifiers or secrets in operational logs, PII-aware redaction and minimal logging reduce privacy risk; collecting full prompts in diagnostic systems is not inherently necessary; this does not directly meet the required outcome to avoid treating all newly flagged legitimate customer requests as genuine safety threats.

B:

For a system whose objective is to show exactly which source version supported a regulated answer, document IDs, timestamps, and citation spans let reviewers reconstruct the evidence used at response time; later source edits should not erase the audit trail; this does not directly meet the required outcome to avoid treating all newly flagged legitimate customer requests as genuine safety threats.

D:

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 avoid treating all newly flagged legitimate customer requests as genuine safety threats.

Question 126

The engineering change request is to enforce per-tool authorization independently. What is the most direct acceptance criterion?

  1. Evaluate precision and recall for the affected safety categories after adjustment
  2. Attempt denied write calls and confirm enforced rejection plus audit records
  3. Red-team malicious instructions inside retrieved pages and record detected unsafe tool intentions
  4. Measure attack success rate and incident severity across protected versions

Explanation

Correct Answer

B:

The accepted change is to enforce per-tool authorization independently; this evaluation checks whether it can stop an agent's broad language permissions from turning into unrestricted business-system privileges, since tools must authorize each action using the caller's allowed scope, regardless of what the LLM requests; natural-language rules are insufficient.

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 enforce per-tool authorization independently works for the current release.

C:

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 enforce per-tool authorization independently works for the current release.

D:

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 enforce per-tool authorization independently works for the current release.

Question 127

A system now includes a component to use documented safety-policy versioning. Which evaluation should accompany the rollout?

  1. Red-team malicious instructions inside retrieved pages and record detected unsafe tool intentions
  2. Replay a labeled policy suite against old and new guardrail versions
  3. Audit that proposer and approver identities are distinct on completed decisions
  4. Induce a circular task and verify deterministic termination and budget reporting

Explanation

Correct Answer

B:

The accepted change is to use documented safety-policy versioning; this evaluation checks whether it can explain changes in moderation behavior across releases and support rollback, since versioned guardrail policies make thresholds, blocklists and exception decisions reproducible in evaluations and incidents.

Incorrect Answers

A:

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 use documented safety-policy versioning works for the current release.

C:

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 use documented safety-policy versioning works for the current release.

D:

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 use documented safety-policy versioning works for the current release.

Question 128

What test should engineers run after they review false positives after moderation updates, rather than relying solely on a successful API response?

  1. Induce a circular task and verify deterministic termination and budget reporting
  2. Reconstruct historical answer citations using retained source references and version records
  3. Attempt denied write calls and confirm enforced rejection plus audit records
  4. Evaluate precision and recall for the affected safety categories after adjustment

Explanation

Correct Answer

D:

The accepted change is to review false positives after moderation updates; this evaluation checks whether it can avoid treating all newly flagged legitimate customer requests as genuine safety threats, since human-reviewed labeled examples distinguish benign business terms from actual prohibited content; thresholds and policy need measured tuning.

Incorrect Answers

A:

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 review false positives after moderation updates works for the current release.

B:

This test concerns the ability to show exactly which source version supported a regulated answer, for which document IDs, timestamps, and citation spans let reviewers reconstruct the evidence used at response time; later source edits should not erase the audit trail; it does not establish whether the change to review false positives after moderation updates works for the current release.

C:

This test concerns the ability to stop an agent's broad language permissions from turning into unrestricted business-system privileges, for which tools must authorize each action using the caller's allowed scope, regardless of what the LLM requests; natural-language rules are insufficient; it does not establish whether the change to review false positives after moderation updates works for the current release.

Question 129

A user-acceptance test fails with the following observation: A chatbot emits disallowed violent guidance even though the retrieval pipeline is healthy. What is the best correction?

  1. Redact sensitive fields at collection and restrict trace access according to retention policy
  2. Put a checked approval state between proposal and execution for payment-changing tools
  3. Audit disputed blocks, label representative examples, and tune supported thresholds carefully
  4. Configure and test the appropriate input and output harm guardrails for the deployed model

Explanation

Correct Answer

D:

The fault signal is that a chatbot emits disallowed violent guidance even though the retrieval pipeline is healthy; the targeted fix works because content filters classify supported harm categories at the inference boundary; thresholds require evaluation for false positives and misses, restoring the ability to limit harmful user prompts and unsafe assistant responses according to an approved safety policy.

Incorrect Answers

A:

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 chatbot emits disallowed violent guidance even though the retrieval pipeline is healthy, which calls for a different repair.

B:

If the failing function were to keep payment or account-changing tool calls under an explicit human authorization step, this change could be relevant because human-in-the-loop approval gates restrict irreversible operations even if the agent's reasoning or external evidence is wrong; the actual symptom is that a chatbot emits disallowed violent guidance even though the retrieval pipeline is healthy, which calls for a different repair.

C:

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 a chatbot emits disallowed violent guidance even though the retrieval pipeline is healthy, which calls for a different repair.

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.

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