Microsoft AI-103 Practice Test 40
Question 390
A production agent has a good final success rate but sometimes violates approval constraints. Which monitor is needed?
- Report the percentage of tasks completed and hide approval-policy denials so success looks higher.
- Keep redacted trace spans with stable identifiers and secured evidence references
- Capture the original tool request, its failure status and correlated model continuation
- Measure unauthorized tool-attempt and policy-denial events alongside task success
Explanation
Correct Answer
D:Outcome success can hide risky intermediate attempts that require security review. Here the relevant capability must let the team monitor agent policy violations independently of success. An alternative targeting identify expensive agent steps solves another problem.
Incorrect Answers
A:Task success is not proof of compliant execution. Instrument unauthorized tool attempts, policy denials and the corresponding oversight outcomes alongside completion metrics.
B:Structured privacy-controlled telemetry balances reproducible investigation with data minimization. The function serves to investigate using privacy-governed traces, whereas the actual requirement is to monitor agent policy violations independently of success.
C:A complete trajectory shows where tool errors turned into unsupported model output. That capability is appropriate for trace tool failure through model continuation, not this separate task of finding how to monitor agent policy violations independently of success.
Question 391
A company needs to keep customer documents out of publicly routable retrieval services. What should be set up?
- Allocate nonoverlapping private CIDRs but leave the search and document-retrieval service's public endpoint unrestricted.
- Use private network connectivity for both search and storage dependencies
- Disable public network access when the resource and service support it
- Validate VNet connectivity, routes, DNS and network security rules end to end
Explanation
Correct Answer
B:A privately reachable model alone does not protect grounding data sent through public retrieval dependencies. In this scenario the team must protect the RAG data path, not only the model. A feature used to avoid ambiguous routing caused by overlapping virtual network CIDRs does not substitute for it.
Incorrect Answers
A:Unique address ranges are a routing prerequisite, not a private-ingress control. Use the supported search/data private endpoint and restrict public access to keep documents off publicly routable retrieval paths.
C:Turning off public ingress closes the unwanted alternative path around private connectivity. That is relevant to deny public ingress to a private AI service, but the present scenario requires an ability to protect the RAG data path, not only the model.
D:Private endpoint reachability depends on both network paths and resolving the correct private address. This helps when a team needs to troubleshoot cross-network private connectivity; it does not establish how to protect the RAG data path, not only the model.
Question 392
An agent intermittently chooses the wrong API function when tools have similar names. What debugging artifact helps?
- Inspect function-choice traces with tool descriptions and schema versions
- Inspect only the final natural-language response and ignore which tool name, version and argument schema the model selected.
- Record each branch's start, end, dependency waits and tool durations
- Measure unauthorized tool-attempt and policy-denial events alongside task success
Explanation
Correct Answer
A:Tool-selection observability reveals ambiguous descriptions and version drift. This directly supports the need to diagnose wrong function selection, unlike a mechanism primarily intended to investigate using privacy-governed traces.
Incorrect Answers
B:The final prose cannot explain a function-selection mismatch. Tool-choice traces and schema versions reveal whether the model called the intended function with valid parameters.
C:Span timelines reveal the true critical path and wasted sequencing. Here the needed outcome is to diagnose wrong function selection; an option designed to profile multi-agent critical-path latency would not directly provide it.
D:Outcome success can hide risky intermediate attempts that require security review. Its primary purpose is to monitor agent policy violations independently of success; the stated constraint instead demands a way to diagnose wrong function selection.
Question 393
A service must reach a private endpoint from a different network segment but traffic fails. What should the network team validate?
- Raise the output-token budget rather than checking private DNS records, VNet routes and security policy.
- Inspect service firewall rules, endpoint association and effective network policy
- Expand a VNet address range while ignoring private endpoint DNS answers and effective routes to the destination subnet.
- Validate VNet connectivity, routes, DNS and network security rules end to end
Explanation
Correct Answer
D:Private endpoint reachability depends on both network paths and resolving the correct private address. This resolves the specific need to troubleshoot cross-network private connectivity, while a measure that helps control outbound traffic to enterprise tool APIs addresses a different requirement.
Incorrect Answers
A:Token allowance is unrelated to private endpoint reachability. Validate name resolution to the private IP and the end-to-end route and firewall.
B:Service network ACLs and private endpoint scopes control allowed clients independently of model quota. That is relevant to diagnose network admission on AI resources, but the present scenario requires an ability to troubleshoot cross-network private connectivity.
C:A wider prefix cannot by itself deliver packets to a private endpoint from another segment. Validate peering, routing/NSGs, endpoint approval and private DNS resolution end to end.
Question 394
An AI workflow's cache serves another user's previously generated answer containing private records. Which cache design fixes this?
- Key every cached answer solely by natural-language question text.
- Allow a cached result to bypass retrieval authorization on subsequent requests.
- Share the same session-cache namespace across every tenant to improve hit rate.
- Key and authorize cached results by tenant, user entitlements, source version and policy state.
Explanation
Correct Answer
D:Caching must never bypass retrieval ACLs; cross-user reuse of protected content is a privacy breach.
Incorrect Answers
A:Two users may ask the same question while having different permissions and private evidence.
B:Cache hits must still respect the current user's entitlements and applicable source restrictions.
C:Cross-tenant reuse can expose another customer's private response and violates isolation.
Question 395
A sensitive agent must make private model calls, retrieve private documents and block accidental public exposure. Which THREE network controls are necessary? Choose THREE.
- Disable public network access when the resource and service support it
- Use private network connectivity for both search and storage dependencies
- Use public search and storage endpoints and assume the model's private endpoint secures every dependency.
- Inspect firewall configurations but deliberately leave public endpoint access enabled for all dependent search and storage resources.
- Renumber all VNets and keep both the model endpoint and retrieval endpoint publicly reachable.
- Use an appropriate private endpoint and private DNS resolution for the service
Explanation
Correct Answers
A:Turning off public ingress closes the unwanted alternative path around private connectivity.
B:A privately reachable model alone does not protect grounding data sent through public retrieval dependencies.
F:Private Link and correct name resolution keep service access on controlled private networking.
Incorrect Answers
C:Private model calls do not secure other services automatically. Each sensitive storage and retrieval dependency needs supported private connectivity and access controls.
D:Observing firewall settings while keeping dependencies publicly exposed does not satisfy the private-only network requirement. Connectivity, private DNS and public-access restrictions must be enforced.
E:Distinct private address ranges do not enforce confidentiality. The three-control objective also requires private endpoint paths, correct private DNS and disabled or restricted public ingress for the model and data services.
Question 396
A multi-agent workflow slows down after adding another specialist. What latency telemetry matters?
- Record each branch's start, end, dependency waits and tool durations
- Compare only the final model response tokens before and after the new specialist is added.
- Inspect function-choice traces with tool descriptions and schema versions
- Correlate retrieved document IDs and ranking scores with agent output quality
Explanation
Correct Answer
A:Span timelines reveal the true critical path and wasted sequencing. The decisive difference is whether the design can profile multi-agent critical-path latency; the competing function serves to identify expensive agent steps.
Incorrect Answers
B:A branch can be slow due to waiting or serial tool dependencies even when final token count is unchanged. Trace each specialist's active time, tool calls and dependency waits.
C:Tool-selection observability reveals ambiguous descriptions and version drift. Its primary purpose is to diagnose wrong function selection; the stated constraint instead demands a way to profile multi-agent critical-path latency.
D:Coupled retrieval and answer telemetry isolates the evidence-selection stage. That is relevant to connect search relevance problems to agent responses, but the present scenario requires an ability to profile multi-agent critical-path latency.
Question 397
A support agent begins retrieving irrelevant knowledge after index changes. What linked signal is useful?
- Log only total monthly model tokens and disregard which retrieved documents influenced failed answers.
- Inspect function-choice traces with tool descriptions and schema versions
- Capture the original tool request, its failure status and correlated model continuation
- Correlate retrieved document IDs and ranking scores with agent output quality
Explanation
Correct Answer
D:Coupled retrieval and answer telemetry isolates the evidence-selection stage. In this scenario the team must connect search relevance problems to agent responses. A feature used to identify expensive agent steps does not substitute for it.
Incorrect Answers
A:Aggregate spend does not explain why retrieval quality changed after index updates. Correlate selected document IDs, ranking scores, source versions and answer-quality results.
B:Tool-selection observability reveals ambiguous descriptions and version drift. Its primary purpose is to diagnose wrong function selection; the stated constraint instead demands a way to connect search relevance problems to agent responses.
C:A complete trajectory shows where tool errors turned into unsupported model output. That is relevant to trace tool failure through model continuation, but the present scenario requires an ability to connect search relevance problems to agent responses.
Question 398
An incident analyst needs to replay an agent decision while excluding private raw customer messages from logs. What logging design fits?
- Retain raw private customer messages in unrestricted debug logs instead of privacy-preserving trace references.
- Record each branch's start, end, dependency waits and tool durations
- Keep redacted trace spans with stable identifiers and secured evidence references
- Capture the original tool request, its failure status and correlated model continuation
Explanation
Correct Answer
C:Structured privacy-controlled telemetry balances reproducible investigation with data minimization. This resolves the specific need to investigate using privacy-governed traces, while a measure that helps monitor agent policy violations independently of success addresses a different requirement.
Incorrect Answers
A:Uncontrolled raw content storage creates unnecessary exposure. Use restricted, redacted spans and stable evidence references sufficient for incident reconstruction.
B:Span timelines reveal the true critical path and wasted sequencing. That is relevant to profile multi-agent critical-path latency, but the present scenario requires an ability to investigate using privacy-governed traces.
D:A complete trajectory shows where tool errors turned into unsupported model output. The function serves to trace tool failure through model continuation, whereas the actual requirement is to investigate using privacy-governed traces.
Question 399
A retrieval cache continues answering from deleted documents long after the search index removes them. What invalidation mechanism matters?
- Increase the cache TTL so deleted passages remain available for longer.
- Invalidate cached excerpts and generated responses when their source revision is retired or deleted.
- Evict only model weights while retaining cached source excerpts.
- Rely on the updated search index alone to erase old cached answer text.
Explanation
Correct Answer
B:Deleting an index entry alone cannot purge all downstream caches that retained its content.
Incorrect Answers
A:A longer TTL aggravates stale evidence and delays removal of superseded source material.
C:Model cache invalidation does not remove source-dependent responses or chunks.
D:Independent caches can still serve deleted content after index removal unless explicitly invalidated.
Question 400
An agent's cost rises and its tool decisions become unsafe despite a stable final-answer score. Which TWO monitoring additions address the hidden regressions? Choose TWO.
- Record only model output text while discarding the identity, cost and authorization outcome of the underlying tool attempts.
- Correlate retrieved document IDs and ranking scores with agent output quality
- Break down prompt and completion tokens by model invocation and agent step
- Assume model tool selection is safe whenever the final-answer quality score stays constant.
- Measure unauthorized tool-attempt and policy-denial events alongside task success
Explanation
Correct Answers
C:Step-level token accounting identifies costly loops and unnecessary context. Token-by-step profiling detects cost growth and policy-attempt telemetry detects unsafe intermediate behavior.
E:Outcome success can hide risky intermediate attempts that require security review.
Incorrect Answers
A:Text-only telemetry cannot expose which agent step drove cost or whether a forbidden tool action was attempted. Record token costs and policy-denial events independently.
B:Coupled retrieval and answer telemetry isolates the evidence-selection stage. That mechanism concerns connect search relevance problems to agent responses, whereas this case tests how to monitor step cost and unsafe tool execution.
D:Final-response ratings can conceal unsafe intermediate tool decisions; they do not substitute for tool-policy telemetry.