McAfee-Secured Website

Exam Code: AI-103

Exam Name: Developing AI Apps and Agents on Azure

Certification Provider: Microsoft

Corresponding Certification: Microsoft Certified: Azure AI Apps and Agents Developer Associate

100% Microsoft AI-103 Guaranteed Success With Testking Training Materials

Get Real AI-103 Exam Questions & Expert Verified Answers as Experienced on the Real Exam!

187 Questions & Answers with Testing Engine

"Developing AI Apps and Agents on Azure Exam", also known as AI-103 exam, is a Microsoft certification exam.

Pass your tests with the always up-to-date AI-103 Exam Engine. Your AI-103 training materials keep you at the head of the pack!

Satisfaction Guaranteed

Money Back Guarantee

Testking provides hassle-free money back guarantee with our products. That is because we have 100% trust in the abilities of our professional and experience product team, and our record is a proof of that.

99.6% PASS RATE
Was: $137.49
Now: $124.99

Microsoft AI-103 Practice Test 14

Question 130

A RAG assistant reads an external page that includes hidden text telling the assistant to export a confidential customer list. The user only asked for a product policy summary. Which defense matches the entry point?

  1. Only check the user’s original prompt and trust every document returned by retrieval.
  2. Treat retrieved page text as a new higher-priority system instruction.
  3. Turn off content-source trace logging and assume the hidden text will be harmless.
  4. Apply document attack detection to untrusted retrieved web pages.

Explanation

Correct Answer

D:

Document shields address untrusted retrieved content; limited tool privileges provide a second barrier to exfiltration.

Incorrect Answers

A:

The malicious instructions entered through an external retrieved document, not the original user message.

B:

A web page is lower-trust task data and must not override the agent’s developer or security rules.

C:

Removing traces offers no protection from attempted tool hijacking inside retrieved content.

Question 131

A finance agent can call high-impact business tools. The security team determines prompt-injection attempts must be stopped rather than merely logged. How should an applicable Prompt Shields control be configured?

  1. Configure Prompt Shields to block detected injection attempts.
  2. Use annotate-only treatment and assume every detected attempt automatically stops execution.
  3. Disable the shield after the first false positive and remove every other authorization check.
  4. Configure the attack detector only to attach risk annotations and continue all tool calls.

Explanation

Correct Answer

A:

A block action prevents detected unsafe inputs from continuing; tests verify the policy is applied at the intended step.

Incorrect Answers

B:

Annotating a detection records it but does not by itself enforce a block.

C:

This eliminates detection and ignores safer options such as carefully tuned exceptions and layered permissions.

D:

Annotate-only treatment does not enforce the requirement to block detected high-risk attacks before privileged actions.

Question 132

A public creative-writing application must reject output that violates its configured safety categories. The model is otherwise permitted to handle benign requests. What provides the relevant guardrail?

  1. Apply only a post-generation text filter after the response has already reached the public interface.
  2. Switch all search queries to exact-match mode and ignore model content.
  3. Enable safety filters for violent or otherwise prohibited outputs.
  4. Let the browser decide whether a model response is safe based on its length.

Explanation

Correct Answer

C:

Content controls evaluate configured unsafe categories and can block or flag input and output according to policy.

Incorrect Answers

A:

Filtering after exposure does not prevent disallowed output from reaching the user.

B:

Retrieval configuration is not an output safety policy for generative responses.

D:

Response length does not establish compliance with content safety requirements.

Question 133

A moderation classifier returns marginal confidence on phrases describing hazardous repairs. Automatic rejection would prevent legitimate support. What is the defensible treatment?

  1. Ignore all moderation classifications when any low-confidence case occurs.
  2. Block all low-confidence requests forever with no appeal or evaluation.
  3. Grant tool execution permissions to an unreviewed request because its score is uncertain.
  4. Route ambiguous low-confidence classifications for policy-guided review or calibrated secondary evaluation.

Explanation

Correct Answer

D:

Uncertainty calls for proportionate review rather than treating an inconclusive score as a certain violation.

Incorrect Answers

A:

One uncertain result does not invalidate high-confidence safety findings.

B:

Uncertain signals alone do not justify indiscriminate rejection of legitimate content.

C:

Ambiguity does not constitute authorization for high-impact operations.

Question 134

A customer chatbot is told 'never reveal account data' in its system prompt. Legal requires a versioned organizational permitted-use policy enforceable outside model text. Which arrangement is sound?

  1. Enforce a separate permitted-use policy beyond the system prompt.
  2. Give every user database admin access but instruct the model not to use it.
  3. Treat the system prompt as the organization's only legal and technical access policy.
  4. Delete the published use policy because a model can infer appropriate behavior.

Explanation

Correct Answer

A:

A documented policy and enforcement layer can be audited and updated without trusting free-form model obedience alone.

Incorrect Answers

B:

Overprivileged tools undermine the policy regardless of conversational instructions.

C:

Text instructions cannot replace external authorization or formal governance requirements.

D:

Inference does not create an accountable organizational permission standard.

Question 135

An autonomous purchasing assistant can place orders that incur charges. Organizational rules require a manager to approve amounts above a limit. What must happen before the action?

  1. Grant the agent unrestricted purchasing rights once it completes an ordinary chat.
  2. Place orders first and email managers afterward only if the amount is high.
  3. Ask the model to simulate a manager's approval in its own reasoning.
  4. Require a verifiable human approval on the specific purchase before the order tool executes.

Explanation

Correct Answer

D:

A confirmed per-action approval enforces the stated authority boundary for financial side effects.

Incorrect Answers

A:

A routine conversation is not authorization to incur organization expenses.

B:

After-the-fact notification does not satisfy a prior-approval policy.

C:

Generated approval text is not independent authorization from the designated human.

Question 136

An IT agent only needs to read deployment health and restart one approved test service. The offered tool schema includes unrestricted shell commands. How should it be narrowed?

  1. Expose narrowly typed read and restart tools instead of unrestricted shell access.
  2. Permit any command if the agent explanation sounds reasonable.
  3. Validate the model's response length but not the actual operation name.
  4. Give the service subscription Owner to avoid permission failures.

Explanation

Correct Answer

A:

Narrow tool contracts reduce accidental and adversarial execution beyond the authorized tasks.

Incorrect Answers

B:

Persuasive model text does not bound executable effects or resource targets.

C:

A well-formed short response may still trigger an unauthorized command.

D:

Broad infrastructure permissions are unnecessary for the two allowed actions.

Question 137

An agent tool calls a payment gateway using an access token. Debug traces currently include entire HTTP headers. What must be changed?

  1. Store complete gateway tokens permanently so reviewers can reproduce every call.
  2. Send access tokens in the agent's final natural-language response for transparency.
  3. Disable gateway authorization during logging to avoid creating token values.
  4. Prevent agents from exposing credential values in tool logs.

Explanation

Correct Answer

D:

Sanitized logs support debugging without exposing reusable authentication secrets.

Incorrect Answers

A:

Retaining raw tokens creates a high-value credential disclosure surface.

B:

User-visible responses should not expose private service credentials.

C:

Bypassing authentication is unsafe and unnecessary for sanitizing diagnostics.

Question 138

A company agent answers questions from HR and Finance documents. Employee access differs, but the model can query one unrestricted search index. What protection is required?

  1. Tell the model to avoid quoting restricted passages after retrieving them.
  2. Filter source documents by effective user permissions.
  3. Disable source metadata because employee identity is sufficient for every document.
  4. Use answer length limits to make leaked documents less likely to appear.

Explanation

Correct Answer

B:

Per-user retrieval authorization prevents confidential content from entering an unauthorized context.

Incorrect Answers

A:

Post-retrieval instruction does not prevent sensitive source exposure to the model.

C:

Document-level entitlement information is needed to enforce different access scopes.

D:

Short output is not a reliable data-access control.

Question 139

A customer sentiment evaluator records user-level ratings linked to conversations. The team wants long-term aggregate trends, not permanent profiles. What retention design fits?

  1. Keep short-lived/aggregate evaluation records and delete personal profiles.
  2. Export individual customer evaluations to public sample prompts for training.
  3. Retain raw personal evaluation notes forever in shared dashboards.
  4. Disable all deletion controls because aggregated analytics were requested.

Explanation

Correct Answer

A:

Data minimization preserves trend analysis without indefinitely retaining identifiable user judgments.

Incorrect Answers

B:

Publishing identifiable examples violates the goal of limited internal analysis.

C:

Unlimited identifiable retention exceeds the stated analytics need and increases exposure.

D:

Aggregation goals do not eliminate retention and deletion obligations for raw records.

Question 140

An analyst bot creates findings from several uploaded compliance reports. Auditors must trace each conclusion to a specific revision of source evidence. What should be recorded?

  1. Record provenance for generated findings linked to source documents.
  2. Keep only the answer text and the model's subjective confidence rating.
  3. Infer citation page numbers from retrieval rank instead of the actual source-document location.
  4. Replace original documents with a generated narrative after processing.

Explanation

Correct Answer

A:

Provenance metadata makes generated claims inspectable against the actual source version.

Incorrect Answers

B:

Confidence without source links cannot establish which evidence supports a finding.

C:

Retrieval position is not a trustworthy page locator or provenance record.

D:

Destroying the source removes the audit trail needed for independent validation.

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