CCAR-P Governance, Safety, and Risk Management Practice Test 5 — 15 Questions

Question 1

A financial services firm runs a Claude-powered loan pre-screening agent that reads applicant emails and drafts a preliminary recommendation. The firm's risk committee insists that no automated decision may be finalized without a documented human review step. Which control most directly satisfies this requirement while preserving the agent's throughput?

Choose an answer to begin — your selection is scored in the full session.

Accepted Answer

Insert a mandatory human-in-the-loop approval gate before any recommendation is written back to the loan origination system.. The risk committee's requirement is about who authorizes a decision, not how the model behaves or how well activity is recorded. Only a human-in-the-loop approval gate makes a person the final authorizing step before the recommendation reaches the loan origination system, yielding the documented review the committee expects while letting the agent handle upstream work automatically.

Answer

Log every model request and response to immutable storage and retain the logs for seven years.

Answer

Increase the model's temperature to zero and add a system prompt instructing Claude to be conservative in its recommendations.

Answer

Enable prompt caching on the applicant-email summarization step to reduce latency and cost per screening.

Question 2

An enterprise architect is designing a Claude deployment where a single prompt may contain data belonging to customers in the EU, Brazil, and California. The legal team requires that each data subject's rights be honored independently and that processing purposes be documented per jurisdiction. Which architectural approach best supports this requirement?

Accepted Answer

Implement data classification and purpose tagging at ingestion so each record carries jurisdiction, lawful basis, and permitted processing purpose metadata that the orchestration layer enforces at prompt assembly time.. Honoring rights independently while documenting purpose per jurisdiction requires machine-readable metadata attached to each record and enforced during prompt assembly. Classification tags for jurisdiction, lawful basis, and permitted purpose let the orchestration layer exclude data whose purpose does not match the current request, which is the only approach here that scales and remains auditable. Centralization, endpoint selection, and manual pruning do not deliver per-subject enforcement or purpose documentation.

Answer

Route all prompts through a single global data lake so that lineage is consistent and easy to report.

Answer

Rely on the model provider's regional endpoints to satisfy all three jurisdictions automatically.

Answer

Ask business users to manually remove data from prompts when it is not needed for the current task.

Question 3

A multinational bank deploys Claude to draft internal policy summaries for staff in the EU and Singapore. Legal requires that personal data embedded in employee questions never leave its region of origin, but the bank wants a single application codebase. Which architecture most directly enforces the residency requirement?

Accepted Answer

Configure regional API endpoints so that EU traffic is served within the EU and Singapore traffic within its region, sharing only stateless application code.. Data residency is a geographic constraint on where processing happens, so the control must shape the request path rather than the contract or the retention window. Routing each jurisdiction to an in-region API endpoint keeps personal data inside its required boundary while a shared, stateless codebase avoids duplicated engineering effort. Redaction and retention settings reduce risk but do not guarantee the data never leaves the region.

Answer

Enable zero data retention on the account so that no request or response content is stored by the provider.

Answer

Deploy the application in one region and rely on the model provider's contractual data processing addendum to cover cross-border transfer.

Answer

Apply client-side redaction of names and identifiers before the request is sent, then send all traffic to a single global endpoint.

Question 4

A financial services firm runs Claude-powered document review for loan applications. The CISO asks the platform team to produce evidence that every model change affecting production was reviewed and approved before deployment. Which governance mechanism most directly satisfies this requirement?

Accepted Answer

Enforce a change-management record that ties each production model or prompt revision to a named approver, its evaluation results, and its deployment timestamp.. Auditors need a traceable link between a deployed revision, the person who approved it, and the evidence that supported the decision. A change-management record that captures approver identity, evaluation results, and deployment time creates that chain, covering both model identifier updates and prompt changes. Logging, failover, and informal notification describe runtime behavior or awareness but cannot demonstrate pre-deployment authorization.

Answer

Publish the current model version in the internal service catalog and notify stakeholders through the monthly engineering newsletter.

Answer

Configure automatic failover to a secondary model identifier whenever the primary model returns elevated error rates.

Answer

Enable verbose request logging on the Anthropic API so every prompt and completion is stored for later inspection by the security team.

Question 5

A multinational insurer wants to quantify how often its Claude-based claims assistant produces outputs that violate its internal tone and fairness policy before it expands the pilot to three new countries. The compliance team needs a repeatable, statistically defensible measurement rather than anecdotal review. Which approach best meets this need?

Accepted Answer

Run a structured evaluation over a stratified sample of representative claims, scoring each output against a written tone-and-fairness rubric with two independent raters.. Quantifying violation frequency requires a defined sample, a consistent scoring instrument, and controls for scorer bias. Stratified sampling over representative claims plus a written rubric scored by two independent raters yields a rate with reliability evidence, which is what lets compliance compare countries and justify or reject expansion on evidence rather than impressions.

Answer

Review the model's system prompt and knowledge sources with legal counsel to confirm the policy language is correctly worded.

Answer

Monitor the assistant's average response latency and error rate in production dashboards during the pilot period.

Answer

Ask the pilot's ten claims adjusters to report any tone or fairness problems they notice in daily use.

Question 6

An insurer uses a Claude-based agent that can call internal tools to look up policy details. During review, the safety team finds that a document uploaded by a claimant contains text instructing the agent to email the full policy database to an external address. The agent has an email tool available. Which control most directly prevents this class of failure?

Accepted Answer

Enforce tool-level authorization so the agent's identity lacks permission to send external email, and constrain tool arguments to validated allowlists.. When untrusted content can reach a model that holds real capabilities, the reliable fix is to remove the dangerous capability from the agent's identity and constrain the arguments of the tools it keeps. Injected text can persuade a model but cannot grant permissions the credential layer denies, so an agent without external-email rights cannot exfiltrate regardless of what the document says. Detection and prompting remain useful layers but are not deterministic.

Answer

Scan uploaded documents with a classifier that flags imperative language and quarantine any document that scores above threshold.

Answer

Add a system prompt instruction telling the model to ignore any instructions found inside uploaded documents.

Answer

Require human approval for every tool invocation the agent proposes, regardless of which tool or argument is involved.

Question 7

A media company uses Claude to moderate user-generated comments at high volume. The risk team wants a control that detects when the moderation model's behavior drifts, for example becoming unusually permissive or aggressive, before it affects the community at scale. Which control best fits this need?

Accepted Answer

Maintain a labeled evaluation set of representative comments and run scheduled regression tests that compare current moderation decisions against expected outcomes, alerting on threshold breaches.. Detecting behavioral drift requires a stable reference: a labeled evaluation set that defines expected moderation outcomes. Running scheduled regression tests against it converts drift into a measurable deviation and supports alerting before community-wide impact. Token monitoring, user appeals, and verdict caching each serve different purposes and cannot reveal systematic changes in how the model judges content.

Answer

Provide an appeal button so users can report comments they believe were moderated incorrectly.

Answer

Monitor average tokens per moderation request to detect changes in comment length.

Answer

Cache moderation verdicts for identical comments to reduce repeated inference costs.

Question 8

A bank is deploying a Claude agent that can call internal tools to move funds between accounts. Risk leadership wants a control that limits the blast radius if the agent is manipulated into performing unauthorized transfers. Which control best addresses this requirement?

Accepted Answer

Enforce authorization and transaction limits in the downstream banking APIs the agent calls, independent of anything the model outputs.. When an agent can take consequential actions, enforcement must live outside the model in the systems that hold authority. Server-side authorization, per-transaction caps, and velocity limits ensure that even a fully manipulated agent cannot exceed policy, because the downstream API rejects anything outside its rules. Prompt instructions, sampling settings, and post-hoc monitoring cannot provide that hard boundary.

Answer

Route all agent traffic through a proxy that logs every tool invocation and alerts the security operations center after transfers complete.

Answer

Add a system prompt instructing the model to never perform transfers above a defined threshold or to accounts not previously seen.

Answer

Increase the model's temperature to zero so its responses become fully deterministic and cannot be manipulated.

Question 9

A hospital network is drafting its AI risk register for a Claude-based discharge-summary assistant. The governance lead wants entries that describe residual risk after existing controls are applied, and that can be assigned an owner and a review cadence. Which TWO characteristics must each risk register entry have to meet this standard? (Choose two.)

Accepted Answer

A residual risk rating that reflects the effect of the controls already in place.. An operational risk register entry must state the exposure that remains after controls and must have someone accountable for watching it on a schedule. Residual rating captures what is actually at stake post-mitigation, while owner and review interval ensure the entry is revisited and acted upon as the assistant and its clinical context change.

Answer

A verbatim copy of the vendor's model card and system prompt.

Answer

A list of every employee who has ever accessed the assistant.

Answer

A projected cost saving attributed to deploying the assistant.

Question 10

A media company wants a lightweight, recurring review of its Claude-powered content moderation assistant. The team has no dedicated compliance staff and wants the cheapest process that still produces defensible evidence of oversight. Which approach fits best?

Accepted Answer

Adopt a scheduled sampling review where a rotating staff member scores a fixed sample of outputs against written criteria and records the results.. Defensible oversight means showing that a named person reviewed outputs against written criteria on a defined schedule and recorded the outcome. Scheduled sampling achieves exactly that with minimal tooling and staffing, and the recorded scores form a durable evidence trail. Vendor documentation and complaint trends are supporting signals, not substitutes for local, documented human review.

Answer

Commission an annual third-party audit with formal attestation and a published report.

Answer

Track the volume of user complaints and treat a stable or declining trend as sufficient evidence of adequate oversight.

Answer

Rely on the vendor's published safety evaluations and model cards as the primary evidence of the assistant's suitability.

Question 11

An insurance company is preparing an AI risk register for its Claude-based claims triage system. The risk team must document controls that reduce the chance of biased or inconsistent decisions affecting policyholders. (Choose two.)

Accepted Answer

Define and version the decision criteria and prompts used for triage, and re-validate them against a representative dataset whenever they change.. Fairness controls require both a defined, versioned decision logic that is re-validated against representative data and a human appeal path that catches harmful outcomes the model produces. Together they create a preventive and a corrective layer. Throughput, token budgets, and model size describe performance or capability, not equity, and cannot substantiate a claim that biased or inconsistent decisions have been controlled.

Answer

Collect aggregate throughput metrics showing how many claims the system processes per hour during peak periods.

Answer

Raise the model's max_tokens setting so the triage system can produce longer, more detailed justifications for each decision.

Answer

Switch to the largest available Claude model on the assumption that greater capability automatically eliminates biased outputs.

Question 12

A media company uses Claude through the Anthropic API to draft articles. Legal counsel asks the platform team to demonstrate that every published draft can be traced to the exact model behavior that produced it, even after Anthropic deprecates older models. The team currently calls the alias claude-sonnet-4-5. Which change best satisfies counsel's requirement?

Accepted Answer

Pin requests to a dated model snapshot identifier and archive the full request parameters and response with each draft.. Traceability to a specific model behavior requires two things: an immutable model identifier and a durable record of the exact input and output. A dated snapshot pins behavior so it can be reproduced or referenced after deprecation, while archiving request parameters and responses preserves the actual interaction. Aliases, caching, and sampling settings modify cost, latency, or variance but none of them establish which model version produced a given published artifact.

Answer

Increase max_tokens and set temperature to zero so outputs become deterministic.

Answer

Log the alias string claude-sonnet-4-5 alongside each draft in the content management system.

Answer

Enable prompt caching so identical requests return consistent outputs across model updates.

Question 13

A public-sector agency must demonstrate to an external auditor that its Claude-based citizen inquiry assistant was operated in line with its approved safety policy throughout the prior fiscal year. The agency has no centralized record of which policy text was in force, when it changed, or who approved each change. Which governance practice should the agency institute first?

Accepted Answer

Establish version-controlled policy documents with recorded approvals and effective dates.. Auditors reconstruct conformance by comparing what happened against the rules that were in force at the time, which requires an authoritative, dated policy history with named approvers. Version-controlled policies with effective dates supply that timeline; without it, no amount of monitoring, testing, or public communication can demonstrate year-long adherence.

Answer

Publish a citizen-facing FAQ describing how the assistant works and what data it uses.

Answer

Deploy an additional monitoring dashboard that tracks assistant uptime and query volume.

Answer

Commission a penetration test of the assistant's public-facing endpoint.

Question 14

A government agency wants assurance that its Claude deployment will not be used to generate content that violates Anthropic's usage policies. Which action most directly supports ongoing policy compliance?

Accepted Answer

Review Anthropic's usage policies and configure the deployment so its use cases fall within permitted categories, documenting that mapping for internal review.. Policy compliance is achieved by understanding the provider's acceptable use categories, deliberately designing the deployment to stay within them, and documenting that alignment so it can be reviewed. Configuration changes, spending tier, and identity obfuscation do not alter the obligation to use the service within policy and provide no evidence of alignment.

Answer

Disable all safety filters so the agency retains maximum control over what the model is allowed to output.

Answer

Route all requests through a third-party gateway that anonymizes the agency's identity from Anthropic.

Answer

Purchase the highest available usage tier so the agency's traffic is treated as a trusted enterprise workload.

Question 15

A financial services firm runs a Claude-powered loan pre-screening agent that reads applicant emails and drafts a preliminary recommendation. The firm's risk committee insists that no automated decision may be finalized without a documented human review step. Which control most directly satisfies this requirement while preserving the agent's throughput?

Choose an answer to begin — your selection is scored in the full session.

Accepted Answer

Enforce the threshold in the tool-execution layer so that any transfer exceeding the limit is rejected unless a human approval token is presented.. When an agent can take consequential action, the authorization boundary must live outside the model, in the component that executes the action. Enforcing the threshold and requiring a human approval token at the tool layer means a manipulated model cannot exceed its authority, because the code performing the transfer refuses. Prompt rules, reviewer models, and sampling adjustments modify model behavior but cannot guarantee that a hijacked agent will decline a prohibited action.

Answer

Lower the agent's temperature and restrict its tool list to a single transfer function with a fixed daily cap.

Answer

Add a system prompt rule stating that transfers above the threshold must be escalated to a human operator.

Answer

Run a second Claude instance as a reviewer that inspects each proposed transfer and vetoes suspicious ones.