Microsoft AI-103 Dumps
Microsoft AI-103 Practice Test 28
Question 270
During a cost review, a data engineer must accept a change for agent scope escalation. The issue to be addressed is: A support agent uses a function tool to change unrelated payroll settings. Which concrete verification action demonstrates that the relevant behavior works?
Explanation
Correct Answer
To verify this change, the education technology company should ask for a forbidden payroll change and inspect denial and audit evidence. Operators still need the implementation to restrict callable tool scope and validate requested operation against user authorization and the preventive requirement to implement allowlisted action schemas and role-based tool access.
Incorrect Answers
For unsafe content output, it can be sensible to submit a categorized safety test set and verify handling by category if a public assistant occasionally produces disallowed harmful instructions. Without the actual change to ask for a forbidden payroll change and inspect denial and audit evidence, the answer lacks the correct operational scope even if the alternate action is technically valid elsewhere.
In the education technology company acceptance review, trying to let an agent invoke any API visible in its service account does not test the exact agent scope escalation condition where a support agent uses a function tool to change unrelated payroll settings; the relevant test is to ask for a forbidden payroll change and inspect denial and audit evidence. Acceptance should instead demonstrate that the service can ask for a forbidden payroll change and inspect denial and audit evidence, using the described failure as the test criterion.
The action to restrict callable tool scope and validate requested operation against user authorization can contribute to agent scope escalation as a implementation step; here, a support agent uses a function tool to change unrelated payroll settings. In the education technology company acceptance review, trying to restrict callable tool scope and validate requested operation against user authorization does not test the exact agent scope escalation condition where a support agent uses a function tool to change unrelated payroll settings; the relevant test is to ask for a forbidden payroll change and inspect denial and audit evidence.
Question 271
At a public library network, the service administrator is preparing a support escalation. The concrete requirement is: A search result tells an agent to disregard instructions and email a secret. Which design action BEST addresses the stated need?
Explanation
Correct Answer
The proposed implementation is to treat retrieval payloads as untrusted data and refuse tool commands embedded in them; this matches the operational requirement that a search result tells an agent to disregard instructions and email a secret. The team can insert a benign-looking malicious document and ensure no unauthorized email tool call. The team should also simulate indirect injection as a required release security test.
Incorrect Answers
Choosing to promote quoted source text to system-level authority cannot directly replace the specified action to treat retrieval payloads as untrusted data and refuse tool commands embedded in them when a search result tells an agent to disregard instructions and email a secret. The actual implementation must treat retrieval payloads as untrusted data and refuse tool commands embedded in them, because the stated service condition is that a search result tells an agent to disregard instructions and email a secret.
The action to insert a benign-looking malicious document and ensure no unauthorized email tool call can contribute to prompt injection from retrieved text as a targeted acceptance check step; here, a search result tells an agent to disregard instructions and email a secret. Choosing to insert a benign-looking malicious document and ensure no unauthorized email tool call cannot directly replace the specified action to treat retrieval payloads as untrusted data and refuse tool commands embedded in them when a search result tells an agent to disregard instructions and email a secret.
Choosing to evaluate safety and fairness across representative languages and user contexts would direct effort toward toxicity and unfair treatment, whose characteristic issue is that a multilingual customer assistant responds unfairly to some user cohorts. The relevant corrective or preventive action here remains to treat retrieval payloads as untrusted data and refuse tool commands embedded in them, with its own validation criteria.
Question 272
After a service audit at a software consultancy, the team observes this failure: A search result tells an agent to disregard instructions and email a secret. Which flawed assumption or implementation choice would MOST DIRECTLY explain why the intended outcome was not achieved?
Explanation
Correct Answer
This design mistake is to promote quoted source text to system-level authority; it conflicts with the actual prompt injection from retrieved text behavior needed when a search result tells an agent to disregard instructions and email a secret. The software consultancy should instead treat retrieval payloads as untrusted data and refuse tool commands embedded in them, followed by a check to insert a benign-looking malicious document and ensure no unauthorized email tool call.
Incorrect Answers
The alternative to simulate indirect injection as a required release security test is a ongoing preventive rule step rather than the faulty assumption being diagnosed. This software consultancy failure-analysis alternative of simulate indirect injection as a required release security test does not diagnose the particular prompt injection from retrieved text defect where a search result tells an agent to disregard instructions and email a secret; the keyed mistake is promote quoted source text to system-level authority.
The action to assume acceptable results in English prove equal performance in all languages addresses toxicity and unfair treatment rather than the decisive prompt injection from retrieved text condition. The software consultancy should not infer that fixing one adjacent AI service validates its unrelated workflow; the requested action is to promote quoted source text to system-level authority.
The alternative to insert a benign-looking malicious document and ensure no unauthorized email tool call is a targeted acceptance check step rather than the faulty assumption being diagnosed. This software consultancy failure-analysis alternative of insert a benign-looking malicious document and ensure no unauthorized email tool call does not diagnose the particular prompt injection from retrieved text defect where a search result tells an agent to disregard instructions and email a secret; the keyed mistake is promote quoted source text to system-level authority.
This approach relates to content filter false positives: a clinical educational assistant blocks harmless health discussions. If the team chose it instead, the targeted check to insert a benign-looking malicious document and ensure no unauthorized email tool call could still fail because the underlying required action is to promote quoted source text to system-level authority.
The alternative to treat retrieval payloads as untrusted data and refuse tool commands embedded in them is a implementation step rather than the faulty assumption being diagnosed. This software consultancy failure-analysis alternative of treat retrieval payloads as untrusted data and refuse tool commands embedded in them does not diagnose the particular prompt injection from retrieved text defect where a search result tells an agent to disregard instructions and email a secret; the keyed mistake is promote quoted source text to system-level authority.
Question 273
The regional hospital is responding to a security tabletop exercise. A site reliability engineer discovers this issue: A search result tells an agent to disregard instructions and email a secret. What corrective step should the team take to restore the intended functionality most directly?
Explanation
Correct Answer
The direct repair is to remove malicious indexed passage and replay the request with hardened boundaries. Once the condition is corrected, the regional hospital should insert a benign-looking malicious document and ensure no unauthorized email tool call. To reduce recurrence, it should also simulate indirect injection as a required release security test; those steps provide different recovery and prevention functions.
Incorrect Answers
This approach relates to false accusation risk: an incident triage agent accuses employees based on unverified telemetry. If the team chose it instead, the targeted check to insert a benign-looking malicious document and ensure no unauthorized email tool call could still fail because the underlying required action is to remove malicious indexed passage and replay the request with hardened boundaries.
During regional hospital remediation, the proposed step to promote quoted source text to system-level authority is not the direct restoration action for prompt injection from retrieved text when a search result tells an agent to disregard instructions and email a secret; the keyed repair is remove malicious indexed passage and replay the request with hardened boundaries. After remediation the team must insert a benign-looking malicious document and ensure no unauthorized email tool call; the action in this option does not itself establish that target recovery outcome.
The action to insert a benign-looking malicious document and ensure no unauthorized email tool call can contribute to prompt injection from retrieved text as a targeted acceptance check step; here, a search result tells an agent to disregard instructions and email a secret. During regional hospital remediation, the proposed step to insert a benign-looking malicious document and ensure no unauthorized email tool call is not the direct restoration action for prompt injection from retrieved text when a search result tells an agent to disregard instructions and email a secret; the keyed repair is remove malicious indexed passage and replay the request with hardened boundaries.
Question 274
A security engineer at a insurance platform wants the following issue to be prevented in later deployments: A search result tells an agent to disregard instructions and email a secret. Which enduring control or engineering guardrail BEST prevents recurrence?
Explanation
Correct Answer
Standardizing the boundary to simulate indirect injection as a required release security test is preventive because the same prompt injection from retrieved text failure could otherwise return: a search result tells an agent to disregard instructions and email a secret. For this insurance platform, the service should also treat retrieval payloads as untrusted data and refuse tool commands embedded in them, and tests should insert a benign-looking malicious document and ensure no unauthorized email tool call so that the ongoing requirement is measurable.
Incorrect Answers
The action to remove malicious indexed passage and replay the request with hardened boundaries can contribute to prompt injection from retrieved text as a corrective remediation step; here, a search result tells an agent to disregard instructions and email a secret. For prevention at the insurance platform, the measure to remove malicious indexed passage and replay the request with hardened boundaries is weaker or aimed at another issue than the prompt injection from retrieved text guardrail to simulate indirect injection as a required release security test needed when a search result tells an agent to disregard instructions and email a secret.
The action to enforce authorized spending scope and idempotency safeguards addresses tool approval boundaries rather than the decisive prompt injection from retrieved text condition. The insurance platform should not infer that fixing one adjacent AI service validates its unrelated workflow; the requested action is to simulate indirect injection as a required release security test.
For prevention at the insurance platform, the measure to promote quoted source text to system-level authority is weaker or aimed at another issue than the prompt injection from retrieved text guardrail to simulate indirect injection as a required release security test needed when a search result tells an agent to disregard instructions and email a secret. The lasting guardrail for this specific threat is to simulate indirect injection as a required release security test, not simply to complete an adjacent service task.
Question 275
During a service audit, a agent designer must accept a change for prompt injection from retrieved text. The issue to be addressed is: A search result tells an agent to disregard instructions and email a secret. Which concrete verification action demonstrates that the relevant behavior works?
Explanation
Correct Answer
The targeted acceptance test is to insert a benign-looking malicious document and ensure no unauthorized email tool call. The design should first treat retrieval payloads as untrusted data and refuse tool commands embedded in them. A robust media platform deployment also needs to simulate indirect injection as a required release security test; passing this check provides relevant evidence but does not certify every possible input or security risk.
Incorrect Answers
The action to simulate indirect injection as a required release security test can contribute to prompt injection from retrieved text as a ongoing preventive rule step; here, a search result tells an agent to disregard instructions and email a secret. In the media platform acceptance review, trying to simulate indirect injection as a required release security test does not test the exact prompt injection from retrieved text condition where a search result tells an agent to disregard instructions and email a secret; the relevant test is to insert a benign-looking malicious document and ensure no unauthorized email tool call.
The action to treat retrieval payloads as untrusted data and refuse tool commands embedded in them can contribute to prompt injection from retrieved text as a implementation step; here, a search result tells an agent to disregard instructions and email a secret. In the media platform acceptance review, trying to treat retrieval payloads as untrusted data and refuse tool commands embedded in them does not test the exact prompt injection from retrieved text condition where a search result tells an agent to disregard instructions and email a secret; the relevant test is to insert a benign-looking malicious document and ensure no unauthorized email tool call.
In the media platform acceptance review, trying to promote quoted source text to system-level authority does not test the exact prompt injection from retrieved text condition where a search result tells an agent to disregard instructions and email a secret; the relevant test is to insert a benign-looking malicious document and ensure no unauthorized email tool call. Acceptance should instead demonstrate that the service can insert a benign-looking malicious document and ensure no unauthorized email tool call, using the described failure as the test criterion.
Choosing to feed ambiguous telemetry and verify the agent refuses unwarranted certainty would direct effort toward false accusation risk, whose characteristic issue is that an incident triage agent accuses employees based on unverified telemetry. The relevant corrective or preventive action here remains to insert a benign-looking malicious document and ensure no unauthorized email tool call, with its own validation criteria.
Question 276
At a education technology company, the responsible-AI reviewer is preparing a release acceptance review. The concrete requirement is: A system gives recommendation scores but no evidence trail for operators. Which design action BEST addresses the stated need?
Explanation
Correct Answer
Choosing to retain explainable decision metadata and cited evidence where supported directly targets the described model decision transparency gap: a system gives recommendation scores but no evidence trail for operators. As a lasting boundary, they also need to require auditable input versions scores and appropriate uncertainty notes, because an implementation by itself does not prove safe operation.
Incorrect Answers
Choosing to claim that generated chain-of-thought text proves the internal decision process cannot directly replace the specified action to retain explainable decision metadata and cited evidence where supported when a system gives recommendation scores but no evidence trail for operators. The actual implementation must retain explainable decision metadata and cited evidence where supported, because the stated service condition is that a system gives recommendation scores but no evidence trail for operators.
The action to ask operators to reproduce a recommendation from stored evidence and policy can contribute to model decision transparency as a targeted acceptance check step; here, a system gives recommendation scores but no evidence trail for operators. Choosing to ask operators to reproduce a recommendation from stored evidence and policy cannot directly replace the specified action to retain explainable decision metadata and cited evidence where supported when a system gives recommendation scores but no evidence trail for operators.
Choosing to treat retrieval payloads as untrusted data and refuse tool commands embedded in them would direct effort toward prompt injection from retrieved text, whose characteristic issue is that a search result tells an agent to disregard instructions and email a secret. The relevant corrective or preventive action here remains to retain explainable decision metadata and cited evidence where supported, with its own validation criteria.
Question 277
After a design meeting at a pharmaceutical research group, the team observes this failure: A system gives recommendation scores but no evidence trail for operators. Which flawed assumption or implementation choice would MOST DIRECTLY explain why the intended outcome was not achieved?
Explanation
Correct Answer
The faulty assumption is claim that generated chain-of-thought text proves the internal decision process. A functioning alternative would be to retain explainable decision metadata and cited evidence where supported, and the relevant validation would be to ask operators to reproduce a recommendation from stored evidence and policy.
Incorrect Answers
The alternative to ask operators to reproduce a recommendation from stored evidence and policy is a targeted acceptance check step rather than the faulty assumption being diagnosed. This pharmaceutical research group failure-analysis alternative of ask operators to reproduce a recommendation from stored evidence and policy does not diagnose the particular model decision transparency defect where a system gives recommendation scores but no evidence trail for operators; the keyed mistake is claim that generated chain-of-thought text proves the internal decision process.
The option to treat a previous model's safety score as permanent for all future versions would normally serve a different safety drift regression requirement, where new model settings cause a previously blocked unsafe request to pass. The alternative may be useful in its own context, but it does not directly supply claim that generated chain-of-thought text proves the internal decision process for the stated decision.
The alternative to retain explainable decision metadata and cited evidence where supported is a implementation step rather than the faulty assumption being diagnosed. This pharmaceutical research group failure-analysis alternative of retain explainable decision metadata and cited evidence where supported does not diagnose the particular model decision transparency defect where a system gives recommendation scores but no evidence trail for operators; the keyed mistake is claim that generated chain-of-thought text proves the internal decision process.
Question 278
The travel support provider is responding to a staged service rollout. A machine-learning developer discovers this issue: A system gives recommendation scores but no evidence trail for operators. What corrective step should the team take to restore the intended functionality most directly?
Explanation
Correct Answer
The appropriate remediation is to capture grounded evidence and approved explanation fields instead of hidden reasoning, because the failure concerns model decision transparency: a system gives recommendation scores but no evidence trail for operators. A meaningful verification should ask operators to reproduce a recommendation from stored evidence and policy, and the durable operating rule should require auditable input versions scores and appropriate uncertainty notes after the immediate fix.
Incorrect Answers
During travel support provider remediation, the proposed step to claim that generated chain-of-thought text proves the internal decision process is not the direct restoration action for model decision transparency when a system gives recommendation scores but no evidence trail for operators; the keyed repair is capture grounded evidence and approved explanation fields instead of hidden reasoning. After remediation the team must ask operators to reproduce a recommendation from stored evidence and policy; the action in this option does not itself establish that target recovery outcome.
The action to ask operators to reproduce a recommendation from stored evidence and policy can contribute to model decision transparency as a targeted acceptance check step; here, a system gives recommendation scores but no evidence trail for operators. During travel support provider remediation, the proposed step to ask operators to reproduce a recommendation from stored evidence and policy is not the direct restoration action for model decision transparency when a system gives recommendation scores but no evidence trail for operators; the keyed repair is capture grounded evidence and approved explanation fields instead of hidden reasoning.
This approach relates to false accusation risk: an incident triage agent accuses employees based on unverified telemetry. If the team chose it instead, the targeted check to ask operators to reproduce a recommendation from stored evidence and policy could still fail because the underlying required action is to capture grounded evidence and approved explanation fields instead of hidden reasoning.
Question 279
A AI application owner at a scientific publisher is conducting a release acceptance review concerning model decision transparency. The observed problem is: A system gives recommendation scores but no evidence trail for operators. Which TWO actions together represent the needed implementation AND a durable preventive safeguard? (Select TWO answers.)
Explanation
Correct Answers
The proposed implementation is to retain explainable decision metadata and cited evidence where supported; this matches the operational requirement that a system gives recommendation scores but no evidence trail for operators. The team can ask operators to reproduce a recommendation from stored evidence and policy. The team should also require auditable input versions scores and appropriate uncertainty notes.
The sustained prevention measure is to require auditable input versions scores and appropriate uncertainty notes. The implementation can still use the separate action to retain explainable decision metadata and cited evidence where supported, but a guardrail constrains future changes. The scientific publisher can measure whether the rule works by arranging to ask operators to reproduce a recommendation from stored evidence and policy.
Incorrect Answers
The action to ask operators to reproduce a recommendation from stored evidence and policy can contribute to model decision transparency as a targeted acceptance check step; here, a system gives recommendation scores but no evidence trail for operators. For prevention at the scientific publisher, the measure to ask operators to reproduce a recommendation from stored evidence and policy is weaker or aimed at another issue than the model decision transparency guardrail to retain explainable decision metadata and cited evidence where supported needed when a system gives recommendation scores but no evidence trail for operators.
For data retention for agents, it can be sensible to enforce scoped memory retention and deletion rules on agent stores if a conversation store preserves customer chats past their approved retention period. Without the actual change to retain explainable decision metadata and cited evidence where supported, the answer lacks the correct operational scope even if the alternate action is technically valid elsewhere.
For prevention at the scientific publisher, the measure to claim that generated chain-of-thought text proves the internal decision process is weaker or aimed at another issue than the model decision transparency guardrail to retain explainable decision metadata and cited evidence where supported needed when a system gives recommendation scores but no evidence trail for operators. The lasting guardrail for this specific threat is to require auditable input versions scores and appropriate uncertainty notes, not simply to complete an adjacent service task.
Question 280
During a design meeting, a quality assurance lead must accept a change for model decision transparency. The issue to be addressed is: A system gives recommendation scores but no evidence trail for operators. Which concrete verification action demonstrates that the relevant behavior works?
Explanation
Correct Answer
To verify this change, the nonprofit organization should ask operators to reproduce a recommendation from stored evidence and policy. Operators still need the implementation to retain explainable decision metadata and cited evidence where supported and the preventive requirement to require auditable input versions scores and appropriate uncertainty notes.
Incorrect Answers
For source attribution controls, it can be sensible to change the handbook and test that old citations cannot silently support new claims if an agent quotes internal handbook statements without showing which version they came from. Without the actual change to ask operators to reproduce a recommendation from stored evidence and policy, the answer lacks the correct operational scope even if the alternate action is technically valid elsewhere.
In the nonprofit organization acceptance review, trying to claim that generated chain-of-thought text proves the internal decision process does not test the exact model decision transparency condition where a system gives recommendation scores but no evidence trail for operators; the relevant test is to ask operators to reproduce a recommendation from stored evidence and policy. Acceptance should instead demonstrate that the service can ask operators to reproduce a recommendation from stored evidence and policy, using the described failure as the test criterion.
The action to retain explainable decision metadata and cited evidence where supported can contribute to model decision transparency as a implementation step; here, a system gives recommendation scores but no evidence trail for operators. In the nonprofit organization acceptance review, trying to retain explainable decision metadata and cited evidence where supported does not test the exact model decision transparency condition where a system gives recommendation scores but no evidence trail for operators; the relevant test is to ask operators to reproduce a recommendation from stored evidence and policy.
Developing AI Apps and Agents on Azure (AI-103) Practice Tests
- Microsoft AI-103 Practice Test 26 - 10 Questions (Last Reviewed 27-Oct-2026)
- Microsoft AI-103 Practice Test 27 - 10 Questions (Last Reviewed 26-Oct-2026)
- Microsoft AI-103 Practice Test 28 - 10 Questions (Last Reviewed 25-Oct-2026)
- Microsoft AI-103 Practice Test 29 - 10 Questions (Last Reviewed 24-Oct-2026)
- Microsoft AI-103 Practice Test 30 - 10 Questions (Last Reviewed 23-Oct-2026)