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 17

Question 160

A user opts out of long-term assistant personalization, but old memory entries remain available in future sessions. What must the service implement?

  1. Keep all memories indefinitely and promise to stop mentioning them.
  2. Move user memory into a shared organization profile so it is no longer personal.
  3. Increase thread expiry but leave the persistent memory database unchanged.
  4. Enforce tenant/user-specific retention and deletion rules on the durable memory store and retrieval.

Explanation

Correct Answer

D:

Opt-out must affect existing persistent records and their access, not just future prompt wording.

Incorrect Answers

A:

Silence in responses does not implement the requested data lifecycle policy.

B:

Renaming the store does not change its privacy or access implications.

C:

Conversation lifetime changes do not remove separately stored durable preferences.

Question 161

An inventory agent should read stock levels and suggest reorder quantities, but its tool credential currently permits editing supplier payment accounts. Which permission is appropriate?

  1. Keep full ERP administrator access and rely on a system prompt to prevent writes.
  2. Enforce least privilege on a tool-backed inventory API.
  3. Allow unrestricted writes until a failed transaction is observed in production.
  4. Give all users supplier-account edit privileges for consistency.

Explanation

Correct Answer

B:

Tool privilege must match the actual business action set rather than the agent's theoretical capabilities.

Incorrect Answers

A:

Prompt rules do not substitute for enforceable API permissions.

C:

Reactive monitoring does not meet least-privilege requirements before execution.

D:

Broadening access increases the damage possible from an agent error or attack.

Question 162

A productivity agent can draft customer messages, but organization policy prohibits sending external email without the user's explicit confirmation. What interface rule applies?

  1. Treat successful generation of a mail draft as approval to submit it automatically.
  2. Send any generated draft automatically as soon as the model says it is ready.
  3. Keep draft creation separate from send execution and require user approval for the specific recipient and content.
  4. Approve all future emails based on one general chat instruction.

Explanation

Correct Answer

C:

A confirmation gate verifies intent and prevents unreviewed outbound communication.

Incorrect Answers

A:

A generated draft is not the user's explicit authorization for the external send operation.

B:

Generated confidence is not user authorization to contact external parties.

D:

A blanket inferred permission does not cover changed recipients or sensitive messages.

Question 163

An agent produces a lengthy self-critique while reviewing a policy answer. The interface needs a concise, auditable explanation but must not expose private internal reasoning. What should it return?

  1. Provide a short evidence-based final rationale and relevant source references without revealing private internal reasoning traces.
  2. Provide only a confidence percentage with no task explanation or evidence.
  3. Publish the agent's raw internal reasoning stream verbatim as proof of reliability.
  4. Delete citations and replace them with a claim that the model thought carefully.

Explanation

Correct Answer

A:

Public justifications should explain decisions from verifiable evidence rather than dump hidden model deliberation.

Incorrect Answers

B:

Confidence alone cannot show why the cited policy supports the response.

C:

Private internal traces are not required evidence and may expose sensitive or misleading content.

D:

Unverifiable assurance does not support the actual answer.

Question 164

An assistant reads product reviews from the web. A review includes a hidden URL instructing the agent to upload private contacts there. Which restriction protects the workflow?

  1. Let the model decide that a URL is approved solely from its hostname wording.
  2. Open every URL in the retrieved content because web search returned it.
  3. Apply generation content filtering but leave external URLs in retrieved snippets directly executable by tools.
  4. Treat retrieved URLs and commands as untrusted data, allowing outbound destinations only through independently approved tool policy.

Explanation

Correct Answer

D:

Retrieval snippets must not authorize data transfer to arbitrary addresses.

Incorrect Answers

A:

Surface-level hostname resemblance cannot replace an allowlist or validation.

B:

A search result is not an authorization source for network side effects.

C:

Content filtering does not independently authorize network destinations or prevent tool-directed injection.

Question 165

An expense agent calls `submit_claim` and the tool requires typed amount, currency and receipt ID. Free-form responses sometimes contain invalid values. What should be enforced?

  1. Give the model direct SQL write access and omit the tool contract.
  2. Enforce typed tool schemas and application-side validation before submitting claims.
  3. Check the response only after the claim is submitted.
  4. Parse the first number from any natural-language reply as the approved amount.

Explanation

Correct Answer

B:

Validated types and required fields prevent malformed model arguments from becoming business transactions.

Incorrect Answers

A:

Unrestricted database writes bypass validation and approval rules.

C:

Post-transaction checking cannot prevent an invalid request from committing.

D:

Incidental numbers may not be the intended authorized claim value.

Question 166

A compliance tool must log the actions an agent attempts but should not let the agent rewrite its own evidence. What separation is necessary?

  1. Disable event recording until an incident is confirmed.
  2. Separate incident-log access from agent tool permissions.
  3. Give the agent delete privileges on its own security logs to clear errors.
  4. Store audit events in editable conversation messages as the only record.

Explanation

Correct Answer

B:

Protecting evidence from the acting principal improves trace integrity and investigation reliability.

Incorrect Answers

A:

Events cannot be reconstructed reliably after the fact if they were never recorded.

C:

Self-deletion allows concealment of unauthorized activity.

D:

Conversation content can be changed or truncated and is not an independent audit log.

Question 167

Security discovers that a connected CRM tool was compromised. Investigators need past execution evidence after cutting access. What should operations do?

  1. Revoke or disable the compromised integration while retaining protected logs and relevant trace evidence.
  2. Leave the tool enabled for several days so investigators can watch more calls.
  3. Erase every log and trace at the moment of disconnection.
  4. Replace only the agent display name while retaining the same active credential.

Explanation

Correct Answer

A:

Containment and evidence preservation can be performed together for a compromised tool path.

Incorrect Answers

B:

Continuing compromised access may increase damage without an approved containment plan.

C:

Deleting records may destroy the only evidence of unauthorized access.

D:

A name change does not revoke a compromised integration token.

Question 168

An agent ignores repeated access-denied results and continues trying higher-privilege operations. Security requires automatic containment. What control is appropriate?

  1. Increase the model temperature so tool attempts become less predictable.
  2. Reset tool traces and allow unlimited future retries.
  3. Grant the agent wider privileges until its attempts stop failing.
  4. Stop unauthorized tool retries and isolate the integration.

Explanation

Correct Answer

D:

An emergency stop limits further access attempts once a dangerous pattern is detected.

Incorrect Answers

A:

Randomness is not a reliable authorization or emergency shutdown mechanism.

B:

Log clearing removes evidence while allowing the harmful loop to continue.

C:

Privilege escalation rewards unauthorized requests instead of containing them.

Question 169

A Python app has an approved Foundry project and a deployed chat model. Developers keep receiving authorization errors because they call an arbitrary public model URL rather than the project endpoint. What is the appropriate configuration step?

  1. Set the model temperature to zero to fix endpoint authorization.
  2. Place a public anonymous URL directly in source code so authentication is bypassed.
  3. Connect with the documented Foundry endpoint and scoped identity.
  4. Send requests to the Azure Storage account endpoint instead of the Foundry model route.

Explanation

Correct Answer

C:

A correct project connection and appropriate credential form the prerequisite for authorized application calls.

Incorrect Answers

A:

Temperature changes response sampling and cannot repair a request made to an unauthorized service endpoint.

B:

Bypassing authentication is not a valid way to connect an enterprise application to an approved project.

D:

A storage endpoint is not a deployed model inference API.

Question 170

A team promotes the same AI app from test to production. The model deployment name differs by environment, but the business logic should remain unchanged. What design minimizes deployment mistakes?

  1. Read deployment names and endpoints from environment configuration.
  2. Hard-code the test model name in the Python source for all environments.
  3. Change the model temperature whenever the environment changes to select another deployment.
  4. Use the Azure subscription display name as the model identifier without resolving the real deployment.

Explanation

Correct Answer

A:

Separating deployment identifiers from application code allows identical tested logic to target the approved environment.

Incorrect Answers

B:

A fixed test deployment name can route production requests to the wrong model or fail outright.

C:

Sampling controls do not choose an endpoint or deployment by name.

D:

A subscription label does not name an inference deployment.

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