AI Safety, Security, and Governance

160. A user-facing assistant must block harmful content in both the user's input and the model's response. Which solution meets these requirements?

1. A Add an instruction to the system prompt telling the model to refuse harmful requests.
2. B Apply a content filter to the model's output only, since the input is user-generated.
3. C Apply an Amazon Bedrock guardrail with content filters configured to evaluate both the input and the output.
4. D Review a sample of conversations each week for harmful content.

161. A generated answer cites a policy clause that does not appear in the retrieved passages. Which solution meets these requirements?

1. A Increase the number of passages retrieved so the model has more context.
2. B Instruct the model in the system prompt to answer only from the provided context.
3. C Apply contextual grounding checks that verify the response is supported by the supplied source material and block or flag ungrounded claims.
4. D Lower the model's temperature so its responses are more deterministic.

162. A defence-in-depth approach must protect a GenAI application rather than relying on one control. Which combination of steps meets these requirements? (Select TWO.)

1. A Publish the application's safety policy on the internal wiki.
2. B Apply a guardrail at the model layer and post-processing validation on the response.
3. C Reduce the maximum output tokens so less content can be generated.
4. D Increase the model's temperature so outputs are harder to predict.
5. E Apply pre-processing filters to the input before it reaches the model.

163. An application must detect attempts to override its instructions through crafted user input. Which combination of steps meets these requirements? (Select TWO.)

1. A Write a longer system prompt instructing the model to reject manipulation.
2. B Remove the system prompt so there is nothing to override.
3. C Scope the application's tool and data permissions so a successful injection reaches nothing consequential.
4. D Apply prompt injection and jailbreak detection to the incoming request.
5. E Raise the model's temperature so injected instructions are less reliably followed.

164. A GenAI application in a VPC must reach the model service without its traffic traversing the internet. Which solution meets these requirements?

1. A Create a gateway VPC endpoint for the model service in the application subnets.
2. B Route the application's traffic through a NAT gateway in a public subnet.
3. C Assign public IP addresses to the application instances and restrict them by security group.
4. D Create an interface VPC endpoint for the model service and restrict it with an endpoint policy.

165. Personal information must not appear in model responses, and it may be present in the retrieval corpus. Which combination of steps meets these requirements? (Select TWO.)

1. A Reduce the number of passages retrieved for each query.
2. B Detect and redact personal information in the corpus before it is indexed.
3. C Apply a guardrail sensitive information policy so any remaining personal data is masked in responses.
4. D Instruct the model in the system prompt not to disclose personal information.
5. E Restrict which principals may query the application.

166. An auditor must establish which data sources contributed to a specific generated response. Which solution meets these requirements?

1. A Record the model identifier used for each invocation.
2. B Record the total number of passages retrieved for each invocation.
3. C Record the timestamp of each invocation and the index version in use.
4. D Record the retrieved source identifiers with each invocation and return them as citations alongside the response.

167. A compliance framework requires documented limitations and intended use for each deployed model configuration. Which solution meets these requirements?

1. A Record the model identifier and its version in the deployment configuration.
2. B Publish the model provider's documentation on the internal wiki.
3. C Produce model cards recording intended use, limitations, and evaluation results, and keep them current with each configuration change.
4. D Record the evaluation metrics in the pipeline's execution history.

168. Users must be able to see why the assistant produced a particular answer. Which solution meets these requirements?

1. A Surface the retrieved sources and, where available, the agent's reasoning trace alongside the answer.
2. B Display the total time taken to produce the answer.
3. C Display the model's confidence score for each answer.
4. D Display the model identifier and the prompt version used.

169. An application's outputs must be assessed for systematic differences in quality across user groups. Which solution meets these requirements?

1. A Ask the model to assess whether its own outputs are fair.
2. B Remove all references to the groups from the evaluation prompts.
3. C Evaluate the outputs on an aggregate dataset and report the overall mean score.
4. D Evaluate the outputs on a dataset stratified by the relevant groups and compare the quality metrics between them.

170. A guardrail blocks a request that a legitimate user needed to make. Which response is appropriate?

1. A Accept the blocks because a guardrail should err toward caution.
2. B Review the blocked examples and adjust the guardrail's configuration, treating false positives as a measurable rate to manage.
3. C Instruct users to rephrase requests until they pass.
4. D Disable the guardrail so legitimate requests are not blocked.

171. An application must verify that a generated answer is supported by the retrieved passages before returning it. Which approach is appropriate?

1. A Apply a grounding check comparing the response against the supplied context, and flag or block unsupported claims.
2. B Check that the response is shorter than the supplied context.
3. C Check that the response cites at least one passage.
4. D Compare the response against the model's previous answers to the same question.

172. An adversarial testing programme must be established for a GenAI application. Which approach is appropriate?

1. A Ask the model to generate adversarial inputs and confirm it resists them.
2. B Rely on the guardrail vendor's testing of their own product.
3. C Test with adversarial inputs once before the initial launch.
4. D Maintain a growing suite of adversarial inputs derived from observed attempts and known technique classes, and run it on every change.

173. An application must ensure that a user can only retrieve documents their role permits. Which approach is appropriate?

1. A Retrieve from the full index and filter the response by role.
2. B Maintain a separate index for each individual user.
3. C Derive the retrieval filter from the authenticated user's role at request time, enforced before retrieval.
4. D Include the user's role in the prompt so the model respects it.

174. Prompts and responses logged for troubleshooting may contain personal information. Which approach is appropriate?

1. A Encrypt the log group and grant the troubleshooting team access.
2. B Mask the sensitive values in the logs and restrict unmasked access to a narrow set of principals.
3. C Reduce the log retention period so the data expires sooner.
4. D Disable logging so the data is never recorded.

175. An organization must demonstrate which data sources contributed to a GenAI system's outputs over a reporting period. Which approach is appropriate?

1. A Record the number of documents in the knowledge base at each month end.
2. B Record the volume of requests handled during the period.
3. C Record which model version was in use during the period.
4. D Record source identifiers with each invocation and retain them for the reporting period.

176. A governance framework must detect when a deployed GenAI system begins violating a policy it previously satisfied. Which approach is appropriate?

1. A Assess compliance thoroughly before each deployment.
2. B Review the system's configuration quarterly.
3. C Require an annual audit of the system's outputs.
4. D Monitor policy-relevant signals continuously and alert on violations, rather than assessing compliance only at deployment.

177. A user must be able to judge how much to rely on a generated answer. Which approach is most useful?

1. A Present the sources the answer drew on so the user can check them directly.
2. B Present the model's name and version.
3. C Present a numeric confidence score alongside the answer.
4. D Present the time taken to generate the answer.

178. An A/B test must establish whether a prompt change affects outputs differently across user groups. Which design is appropriate?

1. A Run the test until the aggregate difference reaches significance.
2. B Run the test on the largest user group and apply the result to all groups.
3. C Stratify the test population by the relevant groups and compare the outcome within each group as well as overall.
4. D Compare the aggregate outcome between the two prompt versions.

179. A custom moderation workflow must handle content categories a managed guardrail does not cover. Which approach is appropriate?

1. A Rely on the managed guardrail alone.
2. B Review all content manually.
3. C Add a Lambda-based classifier for the custom categories in a Step Functions workflow alongside the managed guardrail.
4. D Ask users not to submit such content.

180. A text-to-SQL feature must guarantee that generated queries cannot modify data. Which control is appropriate?

1. A Trust the model to avoid modifications.
2. B Execute generated queries with a read-only database role so any modification statement fails regardless of what the model generates.
3. C Instruct the model to generate only SELECT statements.
4. D Review each generated query manually.

181. A structured output must be enforced so a downstream parser never receives malformed data. Which approach is appropriate?

1. A Wrap the parser in a try-catch.
2. B Instruct the model to produce valid JSON.
3. C Reduce the temperature.
4. D Validate the model's output against a JSON Schema and reject or retry on failure before passing it downstream.

182. A GenAI application's data access must be auditable so a reviewer can see which data each request touched. Which approach is appropriate?

1. A Rely on the data store's own access logs without correlation.
2. B Log only the request count.
3. C Log only failed requests.
4. D Log the data sources and identifiers accessed per request, correlated with the request identifier, and retain the logs.

183. A retention policy must ensure that logged prompts containing personal data are deleted after 90 days. Which approach is appropriate?

1. A Delete the logs manually each quarter.
2. B Apply an S3 lifecycle rule expiring the log objects at 90 days, and set matching retention on any CloudWatch log groups.
3. C Encrypt the logs instead of deleting them.
4. D Retain logs indefinitely for troubleshooting.

184. A regulatory framework requires a decision log for every automated decision a GenAI system makes. Which implementation is appropriate?

1. A Record each decision with its inputs, the model and prompt version, the output, and the timestamp in a durable log.
2. B Record decisions that were disputed.
3. C Record a daily summary count.
4. D Record only the decision outcome.

185. A GenAI system's compliance posture must be monitored for policy violations that emerge after deployment. Which approach is appropriate?

1. A Assess compliance at deployment only.
2. B Rely on user complaints.
3. C Audit the system annually.
4. D Run automated detection for misuse, drift, and policy violations continuously, with alerting and remediation workflows.

186. An agent's reasoning must be surfaced to users so they can understand how it reached a recommendation. Which capability provides this?

1. A The number of tools the agent used.
2. B The agent's total execution time.
3. C Agent tracing that records each reasoning step and tool call, rendered as an explanation alongside the result.
4. D The model's confidence score.

187. A policy-compliant system must document each foundation model's known limitations for reviewers. Which artifact is appropriate?

1. A A model card recording intended use, limitations, evaluation results, and known failure modes.
2. B The application's user guide.
3. C A list of the model's parameters.
4. D The model's API documentation.

188. A guardrail must block a category of content the managed filters do not cover. Which approach is appropriate?

1. A Instruct the model to avoid the category.
2. B Enable every available managed filter at the strictest setting.
3. C Add a custom filter for that category and evaluate its false positive and false negative rates before enabling it.
4. D Review outputs manually for the category.

189. A GenAI application must prevent a user from extracting content they are not entitled to see through careful prompting. Which approach is appropriate?

1. A Log attempts to access restricted content.
2. B Filter the response for restricted content.
3. C Enforce entitlement at retrieval so unauthorized content never enters the context, rather than relying on the model to withhold it.
4. D Instruct the model not to reveal restricted content.

190. A safety control must be tested against inputs designed to bypass it. Which approach is appropriate?

1. A Maintain an adversarial test suite covering known bypass techniques and run it on every change.
2. B Test the control once when it is implemented.
3. C Monitor production for successful bypasses.
4. D Rely on the control vendor's testing.

191. A GenAI application must handle content that is legitimate in one context and inappropriate in another. Which approach is appropriate?

1. A Ask the user to confirm the context.
2. B Block the content in all contexts.
3. C Evaluate the content against the context of the request rather than applying a single rule to all requests.
4. D Allow the content in all contexts.

192. A GenAI system's safety controls must not be bypassed by calling the model directly. Which approach is appropriate?

1. A Apply the controls in the client application.
2. B Monitor for direct model invocations.
3. C Enforce the controls at the service boundary and restrict direct model access so every path is covered.
4. D Document that direct access is prohibited.

193. Data sent to a foundation model must not be used to train the provider's models. Which approach is appropriate?

1. A Send the data only from a private network.
2. B Confirm the service's data handling terms and configure any available opt-out before sending production data.
3. C Remove identifiers from the data.
4. D Encrypt the data before sending it.

194. A GenAI application must isolate one tenant's data from another's in its vector store. Which approach is appropriate?

1. A Filter results by tenant after retrieval.
2. B Enforce tenant separation at retrieval using the authenticated tenant identity, with separate indexes where the isolation requirement is strict.
3. C Include the tenant identifier in the prompt.
4. D Rely on embedding similarity to separate tenants.

195. Prompts containing sensitive data must be retained for troubleshooting without exposing the data. Which approach is appropriate?

1. A Encrypt the log group.
2. B Disable prompt logging.
3. C Reduce the log retention period.
4. D Mask the sensitive fields before the prompt is logged and restrict access to any unmasked copy.

196. A GenAI system must satisfy an audit requirement to show what influenced each decision. Which approach is appropriate?

1. A Record the prompt version, retrieved sources, model version, and response for each decision, correlated by request.
2. B Record the response only.
3. C Record the decisions that were escalated.
4. D Record a daily summary of decisions.

197. An organization must demonstrate that its GenAI system complies with an internal policy over time. Which approach is appropriate?

1. A Document the policy and circulate it.
2. B Assess compliance at deployment.
3. C Evaluate the system against the policy continuously and retain the results as evidence.
4. D Review the system annually.

198. A GenAI system's governance must define who may change its prompts and guardrails. Which approach is appropriate?

1. A Restrict changes to a single individual.
2. B Review changes at the monthly meeting.
3. C Allow any developer to change them.
4. D Restrict those changes to named roles and require review, with every change recorded.

199. A GenAI system processing data from several jurisdictions must meet each one's requirements. Which approach is appropriate?

1. A Identify the requirements applying to each data flow and apply the controls that satisfy each.
2. B Apply the requirements of the strictest jurisdiction to everything.
3. C Process all data in one jurisdiction.
4. D Apply the requirements of the organization's home jurisdiction.

200. A GenAI system must support a request to delete an individual's personal data. Which approach is appropriate?

1. A Identify every store holding the data, including the vector index and logs, and remove it from each.
2. B Delete the data from the primary database.
3. C Mark the data as deleted without removing it.
4. D Wait for the data's retention period to expire.

201. A GenAI system's risk classification must determine how much oversight it receives. Which approach is appropriate?

1. A Classify by the cost of running the system.
2. B Classify by the number of users the system serves.
3. C Apply the same oversight to every system.
4. D Classify by the consequence of an incorrect output, the sensitivity of the data, and the system's autonomy, and apply oversight proportionally.

202. A GenAI application must disclose its limitations to users appropriately. Which approach is appropriate?

1. A Publish the limitations in the terms of service.
2. B State what the system is not reliable for, at the point where a user might rely on it.
3. C Rely on users inferring the limitations.
4. D Describe the limitations in the technical documentation.

203. A GenAI system's fairness must be assessed across user groups. Which approach is appropriate?

1. A Report the aggregate quality score.
2. B Assume the model is fair because it was trained on diverse data.
3. C Evaluate the largest group and generalize.
4. D Evaluate output quality separately for each relevant group and compare, rather than reporting an aggregate.

204. A GenAI system's decisions must be explainable to an affected individual. Which approach is appropriate?

1. A Provide the system's overall accuracy.
2. B Provide the model's confidence score.
3. C Provide the model's name and version.
4. D Provide the information the system used and the reasoning path in terms the individual can evaluate.