Microsoft AI-103 Practice Test 20
Question 190
A tool connector returns an unexpected object type to the agent and it treats the data as trusted instructions. What integration check is needed?
- Provide an explicit supported noninteractive workload-identity credential chain in CI
- Accept every syntactically valid connector object as new agent instructions if its JSON parser tolerates extra fields.
- Treat every connector payload as trusted instructions merely because a supported SDK successfully deserializes it.
- Validate tool-result types and treat returned content as untrusted evidence
Explanation
Correct Answer
D:Even a trusted API can return malformed or attacker-controlled content; validation precedes prompting. Here the relevant capability must let the team enforce typed untrusted tool-result handling. An alternative targeting authenticate Foundry SDK within CI solves another problem.
Incorrect Answers
A:Developer sign-in state is not automatically available to a pipeline; CI needs a supported identity. This helps when a team needs to authenticate Foundry SDK within CI; it does not establish how to enforce typed untrusted tool-result handling.
B:Schema acceptance is not an instruction trust boundary. Validate the tool response against its declared contract and keep external content as untrusted data rather than treating its text as agent directives.
C:Syntactic parsing is not authorization or trust validation. Check expected types, sanitize untrusted content and prevent tool responses from becoming privileged instructions.
Question 191
A pilot model performs poorly only after production rollout. What rollback mechanism should have been prepared?
- Make the deployment identifier an explicit environment configuration value
- Keep a tested previous deployment and a controlled traffic-switch procedure
- Request more regional model quota only after the failed rollout instead of keeping a known-good deployment available.
- Version the agent tool contract and provide a compatibility migration window
Explanation
Correct Answer
B:A known-good model deployment with a tested switch limits recovery time when regressions appear. In this scenario the team must restore serving behavior after a model regression. A feature used to route each environment to its own model deployment does not substitute for it.
Incorrect Answers
A:Model deployment names route requests to intended serving configurations and should not be hard-coded incorrectly. Here the needed outcome is to restore serving behavior after a model regression; an option designed to route each environment to its own model deployment would not directly provide it.
C:Throughput allocation cannot restore behavior of a model that failed quality checks. Preserve a rollback-capable stable deployment and test release switching.
D:Contract versioning allows clients and services to transition without abruptly invalidating old requests. This helps when a team needs to roll out breaking agent interface changes safely; it does not establish how to restore serving behavior after a model regression.
Question 192
A Python production app needs efficient model streaming, noninteractive CI authentication and robust connector parsing. Which THREE implementation choices address these needs? Choose THREE.
- Parse documented required fields and tolerate compatible optional additions
- Deserialize arbitrary connector payloads without enforcing types or trust boundaries.
- Consume a supported streaming inference response and assemble events in order
- Depend on the SDK alone while omitting CI workload credentials and connector-schema validation.
- Store raw connector output verbatim in logs before verifying data sensitivity or schema.
- Provide an explicit supported noninteractive workload-identity credential chain in CI
Explanation
Correct Answers
A:Defensive schema handling prevents harmless additive metadata from breaking integrations.
C:Streaming emits incremental content that a client can progressively display while handling completion and errors.
F:Developer sign-in state is not automatically available to a pipeline; CI needs a supported identity.
Incorrect Answers
B:Blind deserialization neither guarantees correct payload types nor prevents untrusted tool content from influencing the application.
D:An SDK does not replace noninteractive authentication or typed result handling. The three requirements include streaming, authorized CI identity and defensive parsing as separate controls.
E:Unvalidated raw connector logging can expose private data and does not make client parsing robust.
Question 193
An AI model endpoint must serve two departments with very different authorization boundaries. What resource access approach should be preferred?
- Keep a tested previous deployment and a controlled traffic-switch procedure
- Version the agent tool contract and provide a compatibility migration window
- Use one shared model deployment with unrestricted credentials for both departments because they have equal quota entitlement.
- Define least-privilege per-department service identities and access scopes
Explanation
Correct Answer
D:Shared serving infrastructure does not imply shared authorization to every model, file or tool. This resolves the specific need to separate departmental serving permissions, while a measure that helps restore serving behavior after a model regression addresses a different requirement.
Incorrect Answers
A:A known-good model deployment with a tested switch limits recovery time when regressions appear. Its primary purpose is to restore serving behavior after a model regression; the stated constraint instead demands a way to separate departmental serving permissions.
B:Contract versioning allows clients and services to transition without abruptly invalidating old requests. That is relevant to roll out breaking agent interface changes safely, but the present scenario requires an ability to separate departmental serving permissions.
C:Quota ownership doesn't impose department-level resource authorization. Separate access controls and credentials by workload and tenant boundaries as required.
Question 194
A planning agent repeats the same failed search indefinitely, consuming the entire execution budget. Which control should stop the loop?
- Allow the agent unlimited searches until it reports success.
- Increase the maximum context window and assume repeated calls will stop.
- Set a maximum tool-iteration budget and a safe failure or escalation path.
- Ignore identical consecutive failed tool responses.
Explanation
Correct Answer
C:Autonomous planning needs finite execution limits; repeating a failed call cannot be considered progress.
Incorrect Answers
A:Unlimited retries cause repeated failed calls and unbounded compute or token cost.
B:A larger prompt budget does not enforce a limit on execution steps.
D:Ignoring failure signatures prevents detecting that the plan is making no progress.
Question 195
A team will move a tested model into production with minimal rollback risk and verifiable configuration parity. Which THREE deployment practices should it adopt? Choose THREE.
- Keep a tested previous deployment and a controlled traffic-switch procedure
- Expose separate versioned deployments and promote through controlled environments
- Change the production deployment in place without retaining the previous model and configuration reference
- Version model identifiers, deployment settings, tool schemas and prompt artifacts together
- Treat a successful quota increase as proof that staging and production use identical model versions, prompts and rollback policies.
- Use a mutable unversioned production deployment alias while discarding the exact prompt, model and tool configuration used in the last working release.
Explanation
Correct Answers
A:A known-good model deployment with a tested switch limits recovery time when regressions appear. Staging the candidate, versioning the full configuration, and retaining a tested traffic rollback path provide reproducible safer rollout.
B:Keeping production and candidate deployments separate supports comparison and rollback.
D:Generation behavior depends on more than code; recording AI configuration helps reproducible rollbacks.
Incorrect Answers
C:An irreversible deployment replacement makes rollback unpredictable. Preserve pinned artifact versions and a staged transition or rollback mechanism.
E:More TPM says nothing about release configuration parity. Compare versioned deployment manifests, validate traffic-switch and rollback plans, and keep independent tests.
F:An alias without immutable configuration lineage cannot prove staging and production parity or restore the prior tested combination. Versioned releases and rollback controls are required.
Question 196
A developer knows a model name but the app has no indication of which Foundry project hosts it. What configuration must be provided?
- Guess the Foundry project from the model's marketing name and send inference requests without an authenticated project connection.
- Use an authorized production managed identity instead of a personal developer session
- Resolve model deployment identifiers from a versioned runtime configuration
- Configure the application's correct Foundry project connection and deployment mapping
Explanation
Correct Answer
D:Applications need the intended project scope and deployment route, not just an arbitrary model family name. The decisive difference is whether the design can resolve the correct Foundry project and deployment; the competing function serves to select an unambiguous tenant/project resource.
Incorrect Answers
A:A model name does not uniquely identify a customer's Foundry project or deployment. The app needs explicit tenant, endpoint, project and deployed-model mapping with appropriate credentials.
B:Unattended services need their own governed principal with revocable scoped access. Here the needed outcome is to resolve the correct Foundry project and deployment; an option designed to separate personal and production credentials would not directly provide it.
C:Externalizing deployment selection supports staged rollout without code changes. Its primary purpose is to switch target model deployment independently from source code; the stated constraint instead demands a way to resolve the correct Foundry project and deployment.
Question 197
A notebook authenticates using the engineer's account while production should run unattended. How should application identity change?
- Use explicit tenant and resource endpoint identifiers rather than ambiguous display names
- Store only nonsecret endpoints in configuration and use keyless token authentication
- Manage search and model connection identifiers as environment-specific configuration
- Use an authorized production managed identity instead of a personal developer session
Explanation
Correct Answer
D:Unattended services need their own governed principal with revocable scoped access. In this scenario the team must separate personal and production credentials. A feature used to select an unambiguous tenant/project resource does not substitute for it.
Incorrect Answers
A:Connection routing should identify the actual authorized tenant and project, not search by name. Here the needed outcome is to separate personal and production credentials; an option designed to select an unambiguous tenant/project resource would not directly provide it.
B:Keyless credentials prevent leaked static secrets from becoming service access paths. Its primary purpose is to keep sensitive access secrets out of configuration; the stated constraint instead demands a way to separate personal and production credentials.
C:Stage and production resources should not be selected by hard-coded stale identifiers. That is relevant to point each environment to correct project dependencies, but the present scenario requires an ability to separate personal and production credentials.
Question 198
An app is moved from staging to production but still queries a staging knowledge index. What settings need separation?
- Resolve model deployment identifiers from a versioned runtime configuration
- Configure the application's correct Foundry project connection and deployment mapping
- Manage search and model connection identifiers as environment-specific configuration
- Check the calling identity's required inference authorization at the deployment scope
Explanation
Correct Answer
C:Stage and production resources should not be selected by hard-coded stale identifiers. This resolves the specific need to point each environment to correct project dependencies, while a measure that helps switch target model deployment independently from source code addresses a different requirement.
Incorrect Answers
A:Externalizing deployment selection supports staged rollout without code changes. Its primary purpose is to switch target model deployment independently from source code; the stated constraint instead demands a way to point each environment to correct project dependencies.
B:Applications need the intended project scope and deployment route, not just an arbitrary model family name. That is relevant to resolve the correct Foundry project and deployment, but the present scenario requires an ability to point each environment to correct project dependencies.
D:Project access does not necessarily imply permission to invoke every model deployment. The function serves to authorize model inference separately from project access, whereas the actual requirement is to point each environment to correct project dependencies.
Question 199
A multi-agent system delegates investigation to two specialists that edit one shared case file simultaneously. What coordination is needed?
- Let both specialists overwrite the shared case file using last-writer-wins updates.
- Use version checks or isolated edits followed by a controlled merge and conflict resolution.
- Merge all edits into one string without preserving document version identifiers.
- Treat the first specialist response as final and discard the second.
Explanation
Correct Answer
B:Concurrent shared-state mutation can overwrite one specialist's findings; orchestration must handle write conflicts explicitly.
Incorrect Answers
A:Last-writer-wins updates can silently overwrite evidence written by another investigator. Coordinate edits through explicit ownership, version checks or a serialized merge with a conflict record before the shared case file changes.
C:Unversioned concatenation cannot resolve conflicting modifications to the same field.
D:Discarding one branch loses valid investigation contributions instead of coordinating them.
Question 200
A multi-environment Foundry app must avoid embedded keys, select the right production project and use an unattended identity. Which THREE settings directly meet those goals? Choose THREE.
- Configure the application's correct Foundry project connection and deployment mapping
- Hard-code the developer's personal interactive model session as the only credential for the unattended production workload.
- Use an authorized production managed identity instead of a personal developer session
- Store only nonsecret endpoints in configuration and use keyless token authentication
- Choose production project and tenant from display-name similarity alone, without validating the actual resource or tenant identity.
- Use one shared global connection string containing an embedded secret for staging and production, without an environment-specific deployment mapping.
Explanation
Correct Answers
A:Applications need the intended project scope and deployment route, not just an arbitrary model family name. Explicit resource mapping, managed identity and token authentication address routing and secure access.
C:Unattended services need their own governed principal with revocable scoped access.
D:Keyless credentials prevent leaked static secrets from becoming service access paths.
Incorrect Answers
B:A developer session is not a noninteractive production identity and cannot provide reliable keyless service authorization after the developer signs out.
E:Display names can overlap across environments. Production connections need their real resource identifiers, scoped authorization and controlled configuration.
F:A shared secret and ambiguous connection violate the separate production context and unattended keyless-authentication requirements.