Microsoft AB-100 Practice Test 36

Question 350

A marketplace has forty active autonomous agents but nobody can identify the owner of a refund-adjustment agent created last year. What governance control should be established?

A governed agent inventory with owner, purpose, approved tools, risk tier, release version and periodic access recertification.
A list of agent display names with no owner or permitted actions.
An assumption that the creator remains accountable after leaving the organization.
A single unrestricted admin credential shared by every agent to reduce inventory overhead.

Explanation

Correct Answer

Accountability depends on identifiable owners and deployed capabilities throughout the agent lifecycle, including offboarding and recertification.

Incorrect Answers

Display names alone cannot identify responsibility or high-risk tool access.

Departed employees cannot provide an enforceable ongoing governance approval.

Shared administrator identity increases risk and defeats attribution rather than simplifying governance safely.

Question 351

A finance model endpoint uses an API key copied into an agent prompt. The prompt and evaluation transcripts are exported to analysts. How should the credential be protected?

Store it in an approved secret store or managed identity path, remove it from prompts, and rotate exposed material.
Share the same key with every tester because evaluation is an internal activity.
Save the key in training metadata so the model can reproduce past transactions.
Mask the key visually in the chat UI but keep the underlying prompt unchanged.

Explanation

Correct Answer

Credentials must be bound outside model instructions and transcripts; previously exposed keys need containment and rotation.

Incorrect Answers

Internal access does not justify broad secret distribution or shared privilege.

Embedding credentials into model provenance risks a persistent information leak.

Display masking does not prevent a model or logs from reading the secret.

Question 352

A product-catalog page retrieved for a retail support agent contains a hidden HTML comment that says to email the entire customer directory to an external address. Which defense matters?

Treat HTML comments as internal system instructions because they are not visible to customers.
Ask the model to mention that it found a hidden comment before executing the tool.
Permit the export if the catalog page domain is on a list of normal knowledge sites.
Treat retrieved page contents as untrusted data and enforce explicit tool allowlists and user authorization before any outbound contact export.

Explanation

Correct Answer

Indirect prompt injection can be embedded in retrieved content, and action permission must remain independent of that content.

Incorrect Answers

Hidden HTML is still lower-trust source text and cannot gain privileged instruction authority.

Warning about an instruction does not make its unauthorized execution safe.

A legitimate domain can host compromised or malicious document content.

Question 353

A construction recruitment assistant ranks applicants for site-supervisor roles. Aggregate accuracy is stable, but older workers receive fewer interview recommendations with comparable qualifications. What responsible-AI review is needed?

Judge fairness solely from whether recommendation messages sound polite.
Rely on the aggregate score because it is unchanged.
Compare error and selection rates across relevant groups at comparable qualifications, investigate the cause and establish remediation and human appeal.
Remove age-related fields from the report without evaluating correlated predictors.

Explanation

Correct Answer

Subgroup outcome analysis and an accountable remedy detect harmful differences hidden within overall scores.

Incorrect Answers

Courteous wording does not establish equitable selection outcomes.

Global accuracy can conceal materially different opportunities for protected or vulnerable groups.

Proxy variables may preserve disparate impact even when explicit attributes disappear.

Question 354

An energy operator restricts AI operations data to an approved national region. A model deployment is regional, but its tool calls invoke a globally routed third-party monitoring service. Which assessment is mandatory?

Remove region names from prompts so the data no longer carries residency requirements.
Assume the monitoring vendor processes data locally because it returns results quickly.
Approve the design because the language-model resource itself is located in the national region.
Map and constrain the complete data path, including remote tools, logs and subprocessors, to approved locations before release.

Explanation

Correct Answer

Residency encompasses all processing and storage destinations, not just the primary model endpoint.

Incorrect Answers

Hiding an origin label does not change regulatory obligations for the actual records.

Latency cannot establish a provider’s data processing geography.

External tool calls can leave the region even when the model is region-pinned.

Question 355

A marketplace buyer can access invoices for their own purchases, but a search-index role lets an agent retrieve every seller’s invoices. How should grounding authorization work?

Enforce per-user source-level document ACLs on retrieval and verify that returned evidence belongs to the caller.
Use one shared index administrator identity and rely on the response prompt not to reveal secrets.
Retrieve all invoices and hide unauthorized document titles in the final UI.
Allow the model to decide whether an invoice seems relevant to the shopper.

Explanation

Correct Answer

Document-level permission trimming must occur before protected content enters the model’s context.

Incorrect Answers

High-privilege shared access bypasses source authorization, even when the prompt promises discretion.

Hiding titles after retrieval does not undo exposure to the model or its tool traces.

Semantic relevance is not the same as purchase-level entitlement.

Question 356

A finance agent’s approval threshold changed overnight. An auditor needs to determine who changed the threshold and which model version used it for a payment. What audit evidence is required?

A list of successful agent conversations without rule or actor identifiers.
A mutable log file that the same agent can silently rewrite after payment completion.
Append-only events linking actor, approved rule version, model release, timestamp and affected payment correlation IDs.
Only the current threshold value shown in an administrative dashboard.

Explanation

Correct Answer

A trustworthy audit trail captures both configuration change provenance and transaction effect with tamper resistance.

Incorrect Answers

Conversations without policy versions lack accountable causal evidence.

Agent-writable logs do not provide sufficient integrity assurance for high-impact financial changes.

The current value cannot prove what a historic payment used or who approved a change.

Question 357

A retail agent requests a large supplier payment. The requestor is the same employee who would approve it. What control prevents self-approval?

Grant the submitting employee both request and approval roles for speed.
Require a separately authenticated authorized approver with enforced segregation of duties before any payment tool commit.
Accept the agent’s text that a supervisor probably agrees.
Send the employee an informational copy of their own request and treat it as signoff.

Explanation

Correct Answer

Separation-of-duties must be enforced by the transaction system with an independent person and verified identity.

Incorrect Answers

Combining requester and approver roles defeats the central risk control.

A generated assertion is not proof of approval by the designated authority.

An informational message is not an independent approval record.

Question 358

A construction estimating agent calls a subcontractor-rate tool that requires project-specific access. The model attempts to pass a different project ID than the signed-in estimator is assigned. What should the tool do?

Allow the operation whenever the assistant says the job is urgent.
Reject the request using the caller’s verified project entitlements, regardless of the project ID in the agent-generated arguments.
Authorize all contractor projects under a single long-lived service account without per-project checks.
Trust the model-provided project ID because it was part of a valid JSON argument.

Explanation

Correct Answer

Structured arguments are data, not authority; the backend must bind resources to an authenticated caller or approved service scope.

Incorrect Answers

Urgency is not an authorization signal.

A blanket service role would let the agent bypass project boundaries.

A syntactically valid ID may reference another project and cause confidential cost exposure.

Question 359

A utility’s agent registry shows several autonomous maintenance agents with no active operations owner after a team reorganization. What should governance require?

Retain all permissions indefinitely because the agents have previously run without incident.
Assign ownership to an autogenerated model name rather than a responsible person or team.
Delete the registry so orphaned agents no longer appear in compliance reports.
Suspend high-risk permissions until a new accountable owner reviews purpose, tools, incidents and lifecycle status.

Explanation

Correct Answer

Autonomous capabilities need live accountability and periodic recertification, particularly for physical operations.

Incorrect Answers

Historical success does not authorize unowned critical-control automation.

A model name cannot assume legal or organizational accountability.

Removing inventory hides the issue instead of stopping risky operation.

Question 360

A marketplace embeds a merchant-service API secret into a custom model evaluation dataset to simplify testing. What is the correct remediation before promoting the model?

Permit deployment if the model cannot quote the exact key in one interactive test.
Change only the secret’s display formatting without replacing the underlying value.
Remove the secret from accessible datasets, rotate it and use a secure runtime credential binding for test and production.
Keep the secret in evaluation text because the dataset is not publicly indexed.

Explanation

Correct Answer

Training and evaluation artifacts can persist or be exported; secrets require controlled access and exposure response.

Incorrect Answers

A single unsuccessful extraction attempt is not evidence that the credential was not exposed.

Formatting changes do not revoke the compromised credential.

Nonpublic storage still creates unnecessary disclosure paths to model builders and artifacts.

Agentic AI Business Solutions Architect (AB-100) Practice Tests

Microsoft AB-100 Practice Test 34 - 10 Questions (Last Reviewed 11-Oct-2026)
Microsoft AB-100 Practice Test 35 - 10 Questions (Last Reviewed 10-Oct-2026)
Microsoft AB-100 Practice Test 36 - 10 Questions (Last Reviewed 09-Oct-2026)
Microsoft AB-100 Practice Test 37 - 10 Questions (Last Reviewed 08-Oct-2026)
Microsoft AB-100 Practice Test 38 - 10 Questions (Last Reviewed 07-Oct-2026)

Pass Your Microsoft AB-100 Exam Easy!

Question 350

Explanation

Correct Answer

Incorrect Answers

Question 351

Explanation

Correct Answer

Incorrect Answers

Question 352

Explanation

Correct Answer

Incorrect Answers

Question 353

Explanation

Correct Answer

Incorrect Answers

Question 354

Explanation

Correct Answer

Incorrect Answers

Question 355

Explanation

Correct Answer

Incorrect Answers

Question 356

Explanation

Correct Answer

Incorrect Answers

Question 357

Explanation

Correct Answer

Incorrect Answers

Question 358

Explanation

Correct Answer

Incorrect Answers

Question 359

Explanation

Correct Answer

Incorrect Answers

Question 360

Explanation

Correct Answer

Incorrect Answers

Agentic AI Business Solutions Architect (AB-100) Practice Tests