Practice Exams:

Microsoft AI-103 Dumps

EXAM: Developing AI Apps and Agents on Azure
Verified By Experts
Premium Bundle
$49.99

AI-103 Premium ETE File


  • Premium File 187 Questions & Answers
Whats Included:
  • Latest Questions
  • 100% Accurate Answers
  • Fast Exam Updates

Microsoft AI-103 Practice Test 59

Question 580

During a design meeting, a cloud developer must accept a change for prompt injection regression. The issue to be addressed is: A new tool integration lets retrieved snippets influence unauthorized actions. Which concrete verification action demonstrates that the relevant behavior works?

  1. Test a retrieved instruction attempting unauthorized export
  2. Include indirect-injection adversarial cases in agent evaluation
  3. Evaluate only friendly queries that contain no adversarial tool content
  4. Ask questions absent from the knowledge store and verify uncertainty

Explanation

Correct Answer

A:

To verify this change, the customer support center should test a retrieved instruction attempting unauthorized export. Operators still need the implementation to include indirect-injection adversarial cases in agent evaluation and the preventive requirement to isolate tool output from instructions and enforce operation allowlists.

Incorrect Answers

B:

The action to include indirect-injection adversarial cases in agent evaluation can contribute to prompt injection regression as a implementation step; here, a new tool integration lets retrieved snippets influence unauthorized actions. In the customer support center acceptance review, trying to include indirect-injection adversarial cases in agent evaluation does not test the exact prompt injection regression condition where a new tool integration lets retrieved snippets influence unauthorized actions; the relevant test is to test a retrieved instruction attempting unauthorized export.

C:

In the customer support center acceptance review, trying to evaluate only friendly queries that contain no adversarial tool content does not test the exact prompt injection regression condition where a new tool integration lets retrieved snippets influence unauthorized actions; the relevant test is to test a retrieved instruction attempting unauthorized export. Acceptance should instead demonstrate that the service can test a retrieved instruction attempting unauthorized export, using the described failure as the test criterion.

D:

For hallucination counterexample tests, it can be sensible to ask questions absent from the knowledge store and verify uncertainty if a new prompt handles known facts but invents missing ones. Without the actual change to test a retrieved instruction attempting unauthorized export, the answer lacks the correct operational scope even if the alternate action is technically valid elsewhere.

Question 581

During a production incident at a pharmaceutical research group, an AI team must address model a versus b routing. The requirement is: A workload has simple tagging calls and complex reasoning questions. Which THREE actions together implement the needed solution, its preventive guardrail, and a targeted acceptance test? (Select THREE answers.)

  1. Route task categories to suitably capable tested models using policy
  2. Compare stratified quality latency and cost after routing
  3. Define explicit improvement criteria and review budget
  4. Monitor routed quality and cost separately per category
  5. Move simple cases to a validated small model and retain complex fallback
  6. Send every short tagging query to the most expensive reasoning model

Explanation

Correct Answers

A:

The proposed implementation is to route task categories to suitably capable tested models using policy; this matches the operational requirement that a workload has simple tagging calls and complex reasoning questions. The team can compare stratified quality latency and cost after routing. The team should also monitor routed quality and cost separately per category.

B:

The targeted acceptance test is to compare stratified quality latency and cost after routing. The design should first route task categories to suitably capable tested models using policy. A robust pharmaceutical research group deployment also needs to monitor routed quality and cost separately per category; passing this check provides relevant evidence but does not certify every possible input or security risk.

D:

The sustained prevention measure is to monitor routed quality and cost separately per category. The implementation can still use the separate action to route task categories to suitably capable tested models using policy, but a guardrail constrains future changes. The pharmaceutical research group can measure whether the rule works by arranging to compare stratified quality latency and cost after routing.

Incorrect Answers

C:

This approach relates to reflection loop budget: a model repeatedly critiques its own answer without improving it. If the team chose it instead, the targeted check to compare stratified quality latency and cost after routing could still fail because the underlying required action is to route task categories to suitably capable tested models using policy.

E:

The action to move simple cases to a validated small model and retain complex fallback can contribute to model a versus b routing as a corrective remediation step; here, a workload has simple tagging calls and complex reasoning questions. Choosing to move simple cases to a validated small model and retain complex fallback cannot directly replace the specified action to route task categories to suitably capable tested models using policy when a workload has simple tagging calls and complex reasoning questions.

F:

Choosing to send every short tagging query to the most expensive reasoning model cannot directly replace the specified action to route task categories to suitably capable tested models using policy when a workload has simple tagging calls and complex reasoning questions. The actual implementation must route task categories to suitably capable tested models using policy, because the stated service condition is that a workload has simple tagging calls and complex reasoning questions.

Question 582

After a data protection inspection at a travel support provider, the team observes this failure: A workload has simple tagging calls and complex reasoning questions. Which flawed assumption or implementation choice would MOST DIRECTLY explain why the intended outcome was not achieved?

  1. Promote majority vote to a guarantee of correctness without source checks
  2. Send every short tagging query to the most expensive reasoning model
  3. Compare stratified quality latency and cost after routing
  4. Route task categories to suitably capable tested models using policy

Explanation

Correct Answer

B:

This design mistake is to send every short tagging query to the most expensive reasoning model; it conflicts with the actual model a versus b routing behavior needed when a workload has simple tagging calls and complex reasoning questions. The travel support provider should instead route task categories to suitably capable tested models using policy, followed by a check to compare stratified quality latency and cost after routing.

Incorrect Answers

A:

The option to promote majority vote to a guarantee of correctness without source checks would normally serve a different multimodel consensus risk requirement, where two models agree on an answer but share the same unsupported source. The alternative may be useful in its own context, but it does not directly supply send every short tagging query to the most expensive reasoning model for the stated decision.

C:

The alternative to compare stratified quality latency and cost after routing is a targeted acceptance check step rather than the faulty assumption being diagnosed. This travel support provider failure-analysis alternative of compare stratified quality latency and cost after routing does not diagnose the particular model a versus b routing defect where a workload has simple tagging calls and complex reasoning questions; the keyed mistake is send every short tagging query to the most expensive reasoning model.

D:

The alternative to route task categories to suitably capable tested models using policy is a implementation step rather than the faulty assumption being diagnosed. This travel support provider failure-analysis alternative of route task categories to suitably capable tested models using policy does not diagnose the particular model a versus b routing defect where a workload has simple tagging calls and complex reasoning questions; the keyed mistake is send every short tagging query to the most expensive reasoning model.

Question 583

The scientific publisher is responding to a quality regression test. A agent designer discovers this issue: A workload has simple tagging calls and complex reasoning questions. What corrective step should the team take to restore the intended functionality most directly?

  1. Monitor routed quality and cost separately per category
  2. Move simple cases to a validated small model and retain complex fallback
  3. Rebuild a faithful checkpoint from the last validated state
  4. Compare stratified quality latency and cost after routing
  5. Send every short tagging query to the most expensive reasoning model

Explanation

Correct Answer

B:

The direct repair is to move simple cases to a validated small model and retain complex fallback. Once the condition is corrected, the scientific publisher should compare stratified quality latency and cost after routing. To reduce recurrence, it should also monitor routed quality and cost separately per category; those steps provide different recovery and prevention functions.

Incorrect Answers

A:

The action to monitor routed quality and cost separately per category can contribute to model a versus b routing as a ongoing preventive rule step; here, a workload has simple tagging calls and complex reasoning questions. During scientific publisher remediation, the proposed step to monitor routed quality and cost separately per category is not the direct restoration action for model a versus b routing when a workload has simple tagging calls and complex reasoning questions; the keyed repair is move simple cases to a validated small model and retain complex fallback.

C:

The action to rebuild a faithful checkpoint from the last validated state addresses context compression rather than the decisive model a versus b routing condition. The scientific publisher should not infer that fixing one adjacent AI service validates its unrelated workflow; the requested action is to move simple cases to a validated small model and retain complex fallback.

D:

The action to compare stratified quality latency and cost after routing can contribute to model a versus b routing as a targeted acceptance check step; here, a workload has simple tagging calls and complex reasoning questions. During scientific publisher remediation, the proposed step to compare stratified quality latency and cost after routing is not the direct restoration action for model a versus b routing when a workload has simple tagging calls and complex reasoning questions; the keyed repair is move simple cases to a validated small model and retain complex fallback.

E:

During scientific publisher remediation, the proposed step to send every short tagging query to the most expensive reasoning model is not the direct restoration action for model a versus b routing when a workload has simple tagging calls and complex reasoning questions; the keyed repair is move simple cases to a validated small model and retain complex fallback. After remediation the team must compare stratified quality latency and cost after routing; the action in this option does not itself establish that target recovery outcome.

Question 584

A responsible-AI reviewer at a nonprofit organization wants the following issue to be prevented in later deployments: A workload has simple tagging calls and complex reasoning questions. Which enduring control or engineering guardrail BEST prevents recurrence?

  1. Move simple cases to a validated small model and retain complex fallback
  2. Maintain fallback source scopes and outage notices
  3. Send every short tagging query to the most expensive reasoning model
  4. Monitor routed quality and cost separately per category

Explanation

Correct Answer

D:

Standardizing the boundary to monitor routed quality and cost separately per category is preventive because the same model a versus b routing failure could otherwise return: a workload has simple tagging calls and complex reasoning questions. For this nonprofit organization, the service should also route task categories to suitably capable tested models using policy, and tests should compare stratified quality latency and cost after routing so that the ongoing requirement is measurable.

Incorrect Answers

A:

The action to move simple cases to a validated small model and retain complex fallback can contribute to model a versus b routing as a corrective remediation step; here, a workload has simple tagging calls and complex reasoning questions. For prevention at the nonprofit organization, the measure to move simple cases to a validated small model and retain complex fallback is weaker or aimed at another issue than the model a versus b routing guardrail to monitor routed quality and cost separately per category needed when a workload has simple tagging calls and complex reasoning questions.

B:

The action to maintain fallback source scopes and outage notices addresses tool-error fallback rather than the decisive model a versus b routing condition. The nonprofit organization should not infer that fixing one adjacent AI service validates its unrelated workflow; the requested action is to monitor routed quality and cost separately per category.

C:

For prevention at the nonprofit organization, the measure to send every short tagging query to the most expensive reasoning model is weaker or aimed at another issue than the model a versus b routing guardrail to monitor routed quality and cost separately per category needed when a workload has simple tagging calls and complex reasoning questions. The lasting guardrail for this specific threat is to monitor routed quality and cost separately per category, not simply to complete an adjacent service task.

Question 585

During a data protection inspection, a solution architect must accept a change for model a versus b routing. The issue to be addressed is: A workload has simple tagging calls and complex reasoning questions. Which concrete verification action demonstrates that the relevant behavior works?

  1. Simulate a tool outage and inspect user-facing uncertainty and logs
  2. Compare stratified quality latency and cost after routing
  3. Route task categories to suitably capable tested models using policy
  4. Send every short tagging query to the most expensive reasoning model

Explanation

Correct Answer

B:

The targeted acceptance test is to compare stratified quality latency and cost after routing. The design should first route task categories to suitably capable tested models using policy. A robust municipal services team deployment also needs to monitor routed quality and cost separately per category; passing this check provides relevant evidence but does not certify every possible input or security risk.

Incorrect Answers

A:

For tool-error fallback, it can be sensible to simulate a tool outage and inspect user-facing uncertainty and logs if a primary knowledge API is unavailable during a customer session. Without the actual change to compare stratified quality latency and cost after routing, the answer lacks the correct operational scope even if the alternate action is technically valid elsewhere.

C:

The action to route task categories to suitably capable tested models using policy can contribute to model a versus b routing as a implementation step; here, a workload has simple tagging calls and complex reasoning questions. In the municipal services team acceptance review, trying to route task categories to suitably capable tested models using policy does not test the exact model a versus b routing condition where a workload has simple tagging calls and complex reasoning questions; the relevant test is to compare stratified quality latency and cost after routing.

D:

In the municipal services team acceptance review, trying to send every short tagging query to the most expensive reasoning model does not test the exact model a versus b routing condition where a workload has simple tagging calls and complex reasoning questions; the relevant test is to compare stratified quality latency and cost after routing. Acceptance should instead demonstrate that the service can compare stratified quality latency and cost after routing, using the described failure as the test criterion.

Question 586

At a customer support center, the machine-learning developer is preparing a operational readiness review. The concrete requirement is: A bursty application receives 429 responses from an inference endpoint. Which design action BEST addresses the stated need?

  1. Retry instantly without limits and amplify throttling pressure
  2. Stress the endpoint and inspect retry rates and eventual failures
  3. Use bounded exponential backoff with jitter and respect server guidance
  4. Stratify evaluation cases by intent difficulty and harm severity

Explanation

Correct Answer

C:

Choosing to use bounded exponential backoff with jitter and respect server guidance directly targets the described retry backoff gap: a bursty application receives 429 responses from an inference endpoint. As a lasting boundary, they also need to enforce maximum retries budgets and error visibility, because an implementation by itself does not prove safe operation.

Incorrect Answers

A:

Choosing to retry instantly without limits and amplify throttling pressure cannot directly replace the specified action to use bounded exponential backoff with jitter and respect server guidance when a bursty application receives 429 responses from an inference endpoint. The actual implementation must use bounded exponential backoff with jitter and respect server guidance, because the stated service condition is that a bursty application receives 429 responses from an inference endpoint.

B:

The action to stress the endpoint and inspect retry rates and eventual failures can contribute to retry backoff as a targeted acceptance check step; here, a bursty application receives 429 responses from an inference endpoint. Choosing to stress the endpoint and inspect retry rates and eventual failures cannot directly replace the specified action to use bounded exponential backoff with jitter and respect server guidance when a bursty application receives 429 responses from an inference endpoint.

D:

Choosing to stratify evaluation cases by intent difficulty and harm severity would direct effort toward evaluation stratification, whose characteristic issue is that an AI app passes common questions but fails rare high-risk intents. The relevant corrective or preventive action here remains to use bounded exponential backoff with jitter and respect server guidance, with its own validation criteria.

Question 587

After a regional deployment at a manufacturing group, the team observes this failure: A bursty application receives 429 responses from an inference endpoint. Which flawed assumption or implementation choice would MOST DIRECTLY explain why the intended outcome was not achieved?

  1. Stress the endpoint and inspect retry rates and eventual failures
  2. Store only a user-facing reply with no operational identifiers
  3. Retry instantly without limits and amplify throttling pressure
  4. Use bounded exponential backoff with jitter and respect server guidance

Explanation

Correct Answer

C:

The faulty assumption is retry instantly without limits and amplify throttling pressure. A functioning alternative would be to use bounded exponential backoff with jitter and respect server guidance, and the relevant validation would be to stress the endpoint and inspect retry rates and eventual failures.

Incorrect Answers

A:

The alternative to stress the endpoint and inspect retry rates and eventual failures is a targeted acceptance check step rather than the faulty assumption being diagnosed. This manufacturing group failure-analysis alternative of stress the endpoint and inspect retry rates and eventual failures does not diagnose the particular retry backoff defect where a bursty application receives 429 responses from an inference endpoint; the keyed mistake is retry instantly without limits and amplify throttling pressure.

B:

The option to store only a user-facing reply with no operational identifiers would normally serve a different production trace correlation requirement, where a user complaint cannot be tied to a specific model and prompt deployment. The alternative may be useful in its own context, but it does not directly supply retry instantly without limits and amplify throttling pressure for the stated decision.

D:

The alternative to use bounded exponential backoff with jitter and respect server guidance is a implementation step rather than the faulty assumption being diagnosed. This manufacturing group failure-analysis alternative of use bounded exponential backoff with jitter and respect server guidance does not diagnose the particular retry backoff defect where a bursty application receives 429 responses from an inference endpoint; the keyed mistake is retry instantly without limits and amplify throttling pressure.

Question 588

The legal operations unit is responding to a architecture migration. A quality assurance lead discovers this issue: A bursty application receives 429 responses from an inference endpoint. What corrective step should the team take to restore the intended functionality most directly?

  1. Stress the endpoint and inspect retry rates and eventual failures
  2. Retry instantly without limits and amplify throttling pressure
  3. Reduce concurrency and replay failed requests with idempotent retry behavior
  4. Rebuild evaluation runs with held-out labels and controlled context

Explanation

Correct Answer

C:

The appropriate remediation is to reduce concurrency and replay failed requests with idempotent retry behavior, because the failure concerns retry backoff: a bursty application receives 429 responses from an inference endpoint. A meaningful verification should stress the endpoint and inspect retry rates and eventual failures, and the durable operating rule should enforce maximum retries budgets and error visibility after the immediate fix.

Incorrect Answers

A:

The action to stress the endpoint and inspect retry rates and eventual failures can contribute to retry backoff as a targeted acceptance check step; here, a bursty application receives 429 responses from an inference endpoint. During legal operations unit remediation, the proposed step to stress the endpoint and inspect retry rates and eventual failures is not the direct restoration action for retry backoff when a bursty application receives 429 responses from an inference endpoint; the keyed repair is reduce concurrency and replay failed requests with idempotent retry behavior.

B:

During legal operations unit remediation, the proposed step to retry instantly without limits and amplify throttling pressure is not the direct restoration action for retry backoff when a bursty application receives 429 responses from an inference endpoint; the keyed repair is reduce concurrency and replay failed requests with idempotent retry behavior. After remediation the team must stress the endpoint and inspect retry rates and eventual failures; the action in this option does not itself establish that target recovery outcome.

D:

This approach relates to evaluation data leakage: a model appears accurate because benchmark answers are embedded in prompts. If the team chose it instead, the targeted check to stress the endpoint and inspect retry rates and eventual failures could still fail because the underlying required action is to reduce concurrency and replay failed requests with idempotent retry behavior.

Question 589

A AI platform engineer at a financial services division is conducting a operational readiness review concerning retry backoff. The observed problem is: A bursty application receives 429 responses from an inference endpoint. Which TWO actions together represent the needed implementation AND a durable preventive safeguard? (Select TWO answers.)

  1. Isolate test labels from the model generation context and audit evaluation inputs
  2. Retry instantly without limits and amplify throttling pressure
  3. Stress the endpoint and inspect retry rates and eventual failures
  4. Enforce maximum retries budgets and error visibility
  5. Use bounded exponential backoff with jitter and respect server guidance

Explanation

Correct Answers

D:

The sustained prevention measure is to enforce maximum retries budgets and error visibility. The implementation can still use the separate action to use bounded exponential backoff with jitter and respect server guidance, but a guardrail constrains future changes. The financial services division can measure whether the rule works by arranging to stress the endpoint and inspect retry rates and eventual failures.

E:

The proposed implementation is to use bounded exponential backoff with jitter and respect server guidance; this matches the operational requirement that a bursty application receives 429 responses from an inference endpoint. The team can stress the endpoint and inspect retry rates and eventual failures. The team should also enforce maximum retries budgets and error visibility.

Incorrect Answers

A:

For evaluation data leakage, it can be sensible to isolate test labels from the model generation context and audit evaluation inputs if a model appears accurate because benchmark answers are embedded in prompts. Without the actual change to use bounded exponential backoff with jitter and respect server guidance, the answer lacks the correct operational scope even if the alternate action is technically valid elsewhere.

B:

For prevention at the financial services division, the measure to retry instantly without limits and amplify throttling pressure is weaker or aimed at another issue than the retry backoff guardrail to enforce maximum retries budgets and error visibility needed when a bursty application receives 429 responses from an inference endpoint. The lasting guardrail for this specific threat is to enforce maximum retries budgets and error visibility, not simply to complete an adjacent service task.

C:

The action to stress the endpoint and inspect retry rates and eventual failures can contribute to retry backoff as a targeted acceptance check step; here, a bursty application receives 429 responses from an inference endpoint. For prevention at the financial services division, the measure to stress the endpoint and inspect retry rates and eventual failures is weaker or aimed at another issue than the retry backoff guardrail to enforce maximum retries budgets and error visibility needed when a bursty application receives 429 responses from an inference endpoint.

Question 590

During a regional deployment, a compliance specialist must accept a change for retry backoff. The issue to be addressed is: A bursty application receives 429 responses from an inference endpoint. Which concrete verification action demonstrates that the relevant behavior works?

  1. Retry instantly without limits and amplify throttling pressure
  2. Test a retrieved instruction attempting unauthorized export
  3. Stress the endpoint and inspect retry rates and eventual failures
  4. Use bounded exponential backoff with jitter and respect server guidance

Explanation

Correct Answer

C:

To verify this change, the logistics operator should stress the endpoint and inspect retry rates and eventual failures. Operators still need the implementation to use bounded exponential backoff with jitter and respect server guidance and the preventive requirement to enforce maximum retries budgets and error visibility.

Incorrect Answers

A:

In the logistics operator acceptance review, trying to retry instantly without limits and amplify throttling pressure does not test the exact retry backoff condition where a bursty application receives 429 responses from an inference endpoint; the relevant test is to stress the endpoint and inspect retry rates and eventual failures. Acceptance should instead demonstrate that the service can stress the endpoint and inspect retry rates and eventual failures, using the described failure as the test criterion.

B:

For prompt injection regression, it can be sensible to test a retrieved instruction attempting unauthorized export if a new tool integration lets retrieved snippets influence unauthorized actions. Without the actual change to stress the endpoint and inspect retry rates and eventual failures, the answer lacks the correct operational scope even if the alternate action is technically valid elsewhere.

D:

The action to use bounded exponential backoff with jitter and respect server guidance can contribute to retry backoff as a implementation step; here, a bursty application receives 429 responses from an inference endpoint. In the logistics operator acceptance review, trying to use bounded exponential backoff with jitter and respect server guidance does not test the exact retry backoff condition where a bursty application receives 429 responses from an inference endpoint; the relevant test is to stress the endpoint and inspect retry rates and eventual failures.

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