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 13

Question 120

An automated service rotates API keys poorly and repeatedly breaks inference. Its host supports managed identity. Which change removes the shared-key lifecycle?

  1. Use managed identity to avoid rotating embedded access keys.
  2. Put a permanent key in a protected model prompt.
  3. Disable authentication for the service during key rotation.
  4. Encode the existing API key with a different alphabet.

Explanation

Correct Answer

A:

Workload identity replaces distributed reusable static keys with managed credential issuance and authorization.

Incorrect Answers

B:

Prompt content is not a secure or auditable credential store.

C:

Unrestricted inference access is not an acceptable reliability workaround.

D:

Reversible encoding does not eliminate static credential rotation.

Question 121

An agent needs one third-party CRM API secret. Developers copied that secret into tracing attributes to debug failures. What should they implement?

  1. Write the secret in a shared spreadsheet accessible to every developer.
  2. Redact credentials from traces and retrieve the CRM secret through Key Vault at runtime.
  3. Stop logging all errors permanently rather than masking sensitive fields.
  4. Include the secret in every prompt because model instructions are private by definition.

Explanation

Correct Answer

B:

Secret storage and least-privilege retrieval protect credentials while sanitized traces retain useful failure context.

Incorrect Answers

A:

Broad sharing expands the blast radius of credential compromise.

C:

Suppressing every error sacrifices observability without establishing sound secret handling.

D:

Prompts can be logged and replayed and are not intended to hold raw credentials.

Question 122

A developer can invoke the deployed model but receives unauthorized documents from a shared index. Which two control planes must not be conflated?

  1. Remove source ACL metadata to simplify search queries.
  2. Rely only on the model's system prompt to hide unauthorized passages after retrieval.
  3. Assume any user authorized for the model may retrieve all indexed HR files.
  4. Separate model deployment RBAC from index document permissions.

Explanation

Correct Answer

D:

Permission to call inference does not grant access to every source document in the retrieval corpus.

Incorrect Answers

A:

Without permission metadata, source-level authorization becomes harder to enforce.

B:

Content has already entered the model context unless trimmed before use.

C:

Inference access is not equivalent to document entitlement.

Question 123

A customer-facing AI app meets uptime objectives but often fails to complete requests within its contract latency. Which alert design is better?

  1. Alert solely on the number of search documents without testing user requests.
  2. Measure only model token costs and treat low expense as a healthy service.
  3. Page a team only when the entire VM or web process is offline.
  4. Design service health alerts that include user-visible SLO failures.

Explanation

Correct Answer

D:

End-to-end service objectives capture functional failures invisible to a basic web-server heartbeat.

Incorrect Answers

A:

Index size alone is not evidence of successful end-user workflows.

B:

Lower spend may reflect failed requests rather than better operation.

C:

Many partial dependency failures leave the web server alive while users cannot complete tasks.

Question 124

A project manager sets a monthly Azure budget alert and assumes requests will stop when the threshold is reached. What clarification is needed?

  1. Model output temperature is automatically reduced when a budget is exceeded.
  2. Validate a cost budget does not automatically throttle workload spend.
  3. Azure Search indexes become read-only as soon as a budget email is sent.
  4. Any budget alert instantly disables the inference deployment until next month.

Explanation

Correct Answer

B:

A cost budget is not automatically a hard service quota or kill switch for running inference.

Incorrect Answers

A:

Generative sampling does not respond to budget alerts unless application logic is explicitly configured.

C:

An email notification does not by itself change search index permissions.

D:

Budget notifications are not equivalent to automatic resource suspension.

Question 125

An internal assistant repeatedly answers the same public holiday schedule question. User-specific payroll questions must never share cached responses. Which caching policy is appropriate?

  1. Disable expiration for all cached policy answers even after official policy changes.
  2. Cache each employee's payroll answer globally under the common question text.
  3. Cache only shared nonpersonalized results with scope and expiry.
  4. Cache the entire prompt including access tokens for replay across tenants.

Explanation

Correct Answer

C:

Selective caching reduces inference cost without exposing individualized or changing confidential data.

Incorrect Answers

A:

Unbounded retention can serve stale answers from superseded documents.

B:

A global cache key can leak private content across users.

D:

Storing credentials and sensitive prompts introduces avoidable security exposure.

Question 126

A public AI endpoint suddenly receives enormous prompts of repeated text, increasing costs without corresponding task completion. What should operations detect?

  1. Increase maximum output tokens to make the abusive requests more expensive.
  2. Discard token usage metrics to avoid storing user text.
  3. Look only for HTTP 500 errors because misuse cannot occur during successful calls.
  4. Alert on anomalous prompt tokens and enforce request-size limits.

Explanation

Correct Answer

D:

Prompt-size anomalies can signal misuse or cost-amplification incidents despite successful transport responses.

Incorrect Answers

A:

Larger output allowances compound rather than control cost exposure.

B:

Token counts can often be monitored without retaining sensitive payload content.

C:

Expensive malicious requests may succeed while consuming excessive tokens.

Question 127

A model router serves simple FAQ traffic and high-impact compliance explanations. The latter requires stricter evaluation and monitoring. How should routing differ?

  1. Remove traceability from sensitive workflows to minimize latency.
  2. Send all compliance actions to the cheapest unvalidated generative endpoint.
  3. Route high-risk requests to a more extensively monitored model.
  4. Use user prompt length as the only criterion for regulatory significance.

Explanation

Correct Answer

C:

Risk-sensitive routing can impose stronger grounding, review and reliability requirements where consequences are higher.

Incorrect Answers

A:

High-risk cases require stronger provenance, not less operational evidence.

B:

Inference savings do not excuse unmet high-impact accuracy requirements.

D:

Short prompts can still trigger high-impact actions; length is not a sufficient risk signal.

Question 128

The agent's deployment dashboard is green, but ingestion jobs failed to extract 15% of new policy PDFs. What should release monitoring verify before declaring success?

  1. Decrease response length to make the incomplete corpus less noticeable.
  2. Trust the green deployment indicator as proof all upstream files were processed.
  3. Remove unindexed documents from the source inventory without recording the failures.
  4. Check ingestion job outcomes and corpus completeness independently of deployment readiness.

Explanation

Correct Answer

D:

Availability of the running app does not prove that underlying retrieval data is current and complete.

Incorrect Answers

A:

Shorter answers cannot replace missing source evidence.

B:

Deployment state and data pipeline success are separate signals.

C:

Hiding incomplete inputs would conceal the real coverage defect.

Question 129

A public chatbot receives a user message claiming to be a system administrator and ordering it to disregard its safety constraints. Which Foundry control targets this direct input attack?

  1. Add a video summarizer to the application output chain.
  2. Enable the user-prompt attack protection in Prompt Shields at the appropriate user input point.
  3. Change the Azure AI Search ranking algorithm so the hostile prompt is the top result.
  4. Grant the chatbot additional API permissions so it can follow the claimed administrator.

Explanation

Correct Answer

B:

Prompt Shields can identify direct jailbreak-style attempts and apply the configured intervention action.

Incorrect Answers

A:

Video summarization does not classify malicious attempts to override conversational instructions.

C:

Search ranking concerns retrieval relevance and does not enforce model instruction trust boundaries.

D:

Expanding privileges would worsen the risk from an untrusted instruction.

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.

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