CCAR-P domain - 14% of the exam

Governance, Safety & Risk Management

Governance, Safety & Risk Management is 14% of the CCAR-P exam. These are the objectives it covers, each with practice questions, with every answer explained.

The domain in numbers

45
Practice questions
5
Objectives
35
Medium
10
Hard

Where people go wrong

  • Treating an instruction to the model as an access control

    The wrong options rely on a firmer prompt, an instruction to keep secrets confidential or a confirmation step. The model still chooses the taxpayer identifier, the account number or whether approval exists, and the runbook secrets still sit in its context, so a persuasive request can succeed. The correct answers enforce the boundary in code: the handler binds the identifier to the session, checks the account against the manager's book, refuses to run without a recorded second approval, or redacts secrets before passages reach the model.

    See it in a sample question
  • Choosing detection where the requirement is prevention

    The wrong options record or review an action after it has taken effect: audit logging, end-of-day reconciliation, snapshots and logs. Each is a useful control, but the disclosure, payment or deletion has already happened, and a released payment cannot be recalled. Where the stem demands prevention, the correct answer puts the check before the action: a handler that refuses to run, an approval of the whole drafted batch before release, or a prompt that shows the exact resources affected before a destructive step.

    See it in a sample question
  • Assuming a control covers every path the content takes

    The wrong options look for a fault in the model, the patterns, a cache or a replica, when the control itself has a gap in its coverage. A redaction filter that runs on the completed response cannot stop streamed chunks already sent to the browser. An input screen that reads typed messages never sees a worksheet attached as a photograph or PDF. An approval of a draft does not cover documents attached after it. A removed close-claim tool leaves an update-claim tool that can still set the status to closed.

    See it in a sample question
  • Fixing a disparity without measuring it

    The wrong options address unequal outcomes with an instruction, a removed field or equal flag rates, and none produces evidence about the quantity at issue. Notes and report length still carry proxies after a field is deleted, and equal flag rates are the wrong target when true eligibility differs between groups. The correct answers measure the disparity directly: the share of eligible patients missed in each group, inspector grades compared with model priority per district, paired disputes that differ only in writing style, or occupation tallies from a fixed prompt set.

    See it in a sample question

Objectives in this domain

What to study

The official documents this domain's practice questions cite most.

Sample questions from this domain

Free sampleGovernance, Safety & Risk Managementmedium

A national tax agency's online assistant answers questions about a citizen's own filings through a lookup tool whose input includes a taxpayer identifier that the model fills in. In a red-team exercise, testers signed in as one citizen and claimed to be acting for a relative, and in 7 of 40 attempts the assistant retrieved the relative's filing. The agency requires that a signed-in citizen can retrieve their own records alone, and the assistant must keep answering filing questions without staff involvement. What should the architect recommend?

  • AStrengthen the system prompt to say the assistant must look up the signed-in citizen's identifier and refuse others
  • BScreen incoming messages with a classifier that blocks requests mentioning a relative, an agent or another person
  • CHave the tool handler take the taxpayer identifier from the authenticated session and ignore any model-supplied value Correct
  • DLog each lookup with both identifiers and have the fraud team review mismatched pairs in a weekly audit report
Attach tenant or account scope to a tool call server-side from the authenticated session, never from an identifier the model supplies. The defect is that the model chooses whose record is fetched, so anything that persuades the model also widens access. Moving the identifier to the handler and binding it to the authenticated session removes that choice from the model entirely, which turns a probabilistic behaviour into an enforced boundary. Prompt wording and input classifiers lower the attack rate but leave the same capability in place, and audit logs only find the breach afterwards.

Why A is wrong: This is tempting because the failures came from the model being talked into a different identifier, and a firmer instruction would reduce that. It is wrong because the model still controls the identifier, so a sufficiently persuasive request can still succeed; an instruction is not an access control.

Why B is wrong: Input screening is a genuine layer and would catch the red team's phrasing. It is wrong because attackers can rephrase around the classifier, it blocks legitimate questions that mention family, and the identifier is still taken from model output.

Why C is correct: This is correct because the record scope is attached server-side from the session, so no wording in the conversation can change which taxpayer is looked up. The assistant keeps answering filing questions automatically, which meets the no-staff constraint.

Why D is wrong: Audit logging is valuable and an assessor would expect it. It is wrong as the answer because it detects a disclosure days after it has happened rather than preventing it, and it leaves the model in control of whose record is fetched.

Free sampleGovernance, Safety & Risk Managementmedium

A US hospital group's clinician-facing assistant answers questions about patient records in a browser portal. A redaction filter in the application layer scans each completed response and replaces any medical record number with a placeholder. A release then changed the portal from waiting for the full response to rendering tokens as they arrive. In an audit of 300 sessions after the release, 9 screen recordings show a full medical record number appearing on the clinician's screen and then being replaced by the placeholder a moment later. The filter's patterns and the model have not changed since before the release. What is the most likely cause?

  • AThe redaction patterns fail to match the record number format, so the identifiers pass through the filter unchanged
  • BThe model is ignoring a system prompt instruction to omit record numbers, which the filter was meant to back up
  • CThe filter runs on the completed response, so streamed chunks reach the browser before the redaction is applied Correct
  • DThe portal caches an earlier unredacted copy of each response and displays it before loading the redacted one
An output filter only protects what it inspects before delivery, so streaming responses need the filter on the streamed path rather than on the completed text. Output filtering is a control enforced in code outside the model, but it is only effective at the point where it sits in the delivery path. When the portal began rendering tokens as they arrived, text started reaching the browser before the full response existed, while the filter still ran on the completed response. The later replacement proves the patterns work; the gap is the ordering of filter and delivery.

Why A is wrong: A pattern gap is the usual first suspect when an identifier leaks through a filter, which makes this tempting. It is wrong because each number was replaced by the placeholder moments later, which proves the patterns do match it; the leak is about when the filter runs, not what it matches.

Why B is wrong: It is tempting to treat any leaked identifier as the model disobeying its instructions. It is wrong because the filter is the control, the model and its behaviour are unchanged, and a model emitting the number is exactly what the filter exists to catch; the filter's placement is what changed.

Why C is correct: The only change was the move to streaming, and the symptom is a number that appears and is then redacted. A filter that inspects the finished response cannot stop chunks that were already sent to the client, so an output control must sit on the streamed path, for example by buffering until a span has been checked.

Why D is wrong: A stale client cache can produce a brief flash of old content, so this sounds like a reasonable explanation. It is wrong because the release changed rendering to streaming rather than caching, and the timing matches tokens arriving ahead of a post-completion filter.

Free sampleGovernance, Safety & Risk Managementmedium

An online retailer uses Claude to read free-text refund disputes and route each one either to automatic approval or to a manual review queue that adds about six days to a refund. Customer complaints suggest that disputes written in non-standard English are sent to manual review more often. The retailer's data minimisation policy forbids collecting customers' first language or ethnicity, and the head of risk wants evidence, before the next release ships, of whether writing style alone changes the routing outcome. What should the architect recommend?

  • AAdd a system prompt instruction telling the model to judge each dispute on its facts and to disregard spelling, grammar and phrasing.
  • BBuild a test set of paired disputes with identical facts written in different English styles, and compare routing rates across each pair. Correct
  • CBegin recording each customer's first language at checkout so that production routing rates can be compared between language groups.
  • DHave the model attach a self-assessed fairness score to every routing decision and send any low-scoring dispute to manual review.
Counterfactual paired tests can measure whether an irrelevant input feature changes model outcomes without collecting protected or personal attributes. Bias testing needs a measurement in which the only thing that varies is the suspected cause. Paired disputes with the same facts but different writing styles isolate style as the variable, so a gap in routing rates is direct evidence of style-driven treatment. Because the pairs are constructed test data, the retailer gets that evidence before release without collecting first language or ethnicity, which keeps it inside its data minimisation policy.

Why A is wrong: This is tempting because it states the desired behaviour directly and costs almost nothing to ship. It is wrong because an instruction is a statement of intent, not evidence: the head of risk asked whether writing style changes the outcome, and nothing here measures that before or after the change.

Why B is correct: Counterfactual pairs hold the facts constant and vary only the writing style, so any difference in routing rate is attributable to style. The test needs no personal data about real customers, so it respects the data minimisation policy, and it runs before release, which is when the evidence is required.

Why C is wrong: Disaggregating outcomes by group is a recognised bias-testing technique, which makes this attractive. It is wrong because the policy forbids collecting first language, and production data would only show the effect after the release had already shipped, not before it.

Why D is wrong: This looks like a built-in safeguard that needs no extra data. It is wrong because a model's self-reported score is not a measurement of its own bias, and routing more disputes to manual review would add delay without showing whether style drives the routing.

Other domains in this exam

See also the CCAR-P cert hub, the study guide, and the cheat sheet.

Examworthy is not affiliated with or endorsed by Anthropic. Original, blueprint-aligned practice material only.