Amazon AWS Certified Generative AI Developer - Professional AIP-C01 Practice Test 32

Question 310

A guardrail standard tier handles code prompt attacks differently. Which enforcement decision best mitigates this failure?

Assume Classic tier always detects prompt attacks hidden in arbitrary code and every language.
Enable Standard tier without reviewing its required cross-Region guardrail destinations.
Validate deployed tier/model feature support rather than assuming universal filters.
Reuse a guardrail from another model/Region without testing its compatible policies.

Explanation

Correct Answer

Validate which safeguard tier, supported language/code protections and cross-Region processing profile are deployed for the actual model and application.

Incorrect Answers

Tier capabilities and language support differ.

Standard-tier routing has processing geography implications.

Tier, model and Region support are not universal.

Question 311

A user input contains malformed markup designed to override XML prompt fields. What should be implemented before letting the model answer or act?

Concatenate raw user XML into a privileged system prompt with delimiter tags taken at face value.
Treat parsed fields as data and construct trusted instructions separately.
Allow a user-provided <system> element to override actual application rules.
Disable XML parser validation because the language model seems able to read broken tags.

Explanation

Correct Answer

Parse user markup as untrusted fields and assemble trusted instructions through real system/API roles and validated escaping.

Incorrect Answers

Malformed user tags can change the apparent instruction hierarchy.

Markup names inside user data have no trusted role authority.

Natural-language interpretation is not a secure parser or schema control.

Question 312

Model responses are moderated, but a connected agent tool can still transfer restricted data to an unauthorized destination. Select TWO independently enforced controls at the tool execution boundary.

Apply scoped IAM permissions to downstream executor.
Treat Bedrock output passing content moderation as permission to send money.
Use the same unrestricted banking administrator for all end users and trust their chat request.
Permit unlimited transfers once the request JSON satisfies the tool schema.
Require business/record-level authorization for each action.

Explanation

Correct Answer

Restrict the downstream action executor to precisely the allowed IAM actions/resources.

Check business actor, record ownership and amount policy independently for each requested transfer.

Incorrect Answers

A safe-text classification is not a financial transaction authorization.

Backend role breadth cannot replace individual transfer entitlements.

Well-typed arguments can still exceed amount limits or account rights.

Question 313

A contextual grounding check flags irrelevant but factually sourced response. Which enforcement decision best mitigates this failure?

Interpret a high groundedness score as proof the reply answers the user's actual question.
Treat low relevance as definitive evidence every cited fact is false.
Handle relevance and grounding scores as separate evaluation signals.
Ignore source/query/response qualifiers and report a composite score without checking policy applicability.

Explanation

Correct Answer

Contextual grounding tests factual support versus source; relevance tests whether the answer addresses the query, so these scores need distinct thresholds and triage.

Incorrect Answers

Text can be factually supported yet irrelevant to the requested topic.

A tangential answer may contain true facts but miss user intent.

Grounding assessments require correctly supplied source and query content.

Question 314

A streaming response reveals unsafe content before guardrail completes. What should be implemented before letting the model answer or act?

Stream every raw token to the client and remove prohibited text only after the guardrail finishes.
Design buffering or early moderation to prevent premature unsafe token exposure.
Switch to asynchronous guardrail mode and assume it masks PII before output is streamed.
Show an immediate successful answer flag after receiving the first safe chunk.

Explanation

Correct Answer

Use synchronous buffered guardrail scanning or equivalent verified early moderation for safety-critical streams; async mode may expose unsafe chunks first.

Incorrect Answers

Content already sent cannot be undisclosed.

AWS documents that asynchronous guardrail mode cannot mask sensitive information.

Later chunks may trigger blocking or generation failure.

Question 315

A production guardrail update changes blocked topics unexpectedly. Which safety control actually addresses the reported weakness?

Version policies and compare allowed/blocked test sets before rollout.
Deploy the revised denied-topic definition to all traffic with no representative regression set.
Record only the model ID while omitting the guardrail version used for a decision.
Treat every increase in block count as proof of stronger safety without checking false positives.

Explanation

Correct Answer

Compare versioned guardrail policies on both allowed and blocked test cases and roll out only after measured false-positive and false-negative acceptance.

Incorrect Answers

An unexpected policy change can overblock or underblock real users.

Without the policy version, enforcement behavior cannot be reconstructed.

More blocking may reflect harmful overblocking of legitimate topics.

Question 316

A private subnet needs to call Bedrock without public internet. Which control is required in this deployment?

Create only an S3 gateway endpoint and assume it carries Bedrock Runtime requests.
Create the supported bedrock-runtime interface VPC endpoint with private DNS/network controls.
Use a bedrock-agent control-plane endpoint for raw Converse inference calls.
Open outbound port 443 to all public addresses through an Internet gateway despite the private-path requirement.

Explanation

Correct Answer

Create the supported bedrock-runtime interface VPC endpoint and appropriate security groups/private DNS so private-subnet callers can invoke inference without NAT.

Incorrect Answers

An S3 endpoint does not expose the Bedrock Runtime service.

Agent management/runtime and foundation-model runtime endpoints are distinct.

A public egress route does not meet the requested PrivateLink access boundary.

Question 317

An API uses the Bedrock control endpoint but only runtime PrivateLink exists. What should be verified for this identity, network or data boundary?

Send ListFoundationModels through the runtime interface endpoint because it already supports Converse.
Configure the corresponding endpoint type for control-plane calls.
Enable model streaming permissions to make an absent control endpoint appear.
Change only the subnet DNS suffix while leaving no bedrock control endpoint provisioned.

Explanation

Correct Answer

The Bedrock control API has a different interface endpoint service (`bedrock`) than `bedrock-runtime`; provision the endpoint actually used by metadata/model operations.

Incorrect Answers

Control-plane actions are not implemented on the runtime API endpoint.

IAM changes cannot create missing PrivateLink network routes.

DNS cannot route a control-plane service through an endpoint that does not exist.

Question 318

A private endpoint resolves public Bedrock IPs. Which security or privacy control closes the described gap?

Verify private DNS and VPC resolver settings from the calling subnet.
Grant the role additional InvokeModel permissions without checking DNS resolution.
Create only a network ACL permitting public addresses and call the path private.
Remove the VPC interface endpoint and add an S3 bucket policy.

Explanation

Correct Answer

Enable private DNS on the actual interface endpoint and verify VPC resolver/subnet DNS routing; otherwise callers may resolve a public Bedrock address.

Incorrect Answers

Permissions do not alter a public hostname's IP routing.

Public DNS resolution still leads to public service ingress.

Object policies cannot fix Bedrock Runtime private DNS.

Question 319

A model role can invoke every available foundation model. Which control is required in this deployment?

Use bedrock:InvokeModel on Resource * for every application indefinitely.
Scope IAM model and inference-profile actions to approved resources where supported.
Permit only ListFoundationModels and assume model discovery authorizes inference.
Check model output for toxicity as the only limit on which models the role may use.

Explanation

Correct Answer

Constrain model invocation to approved model/inference-profile ARNs and supported actions, respecting necessary cross-Region destination permissions.

Incorrect Answers

Wildcard invocation permits unrelated models and can exceed governance scope.

Metadata read permission does not grant generation.

Content policy is not IAM resource scoping.

Question 320

An enterprise wants to approve an inference profile for use across accounts before deployment, but organization-level Region restrictions might block one or more required profile destinations. Which preflight governance check prevents future AccessDenied failures?

Allow only the source Region in the SCP and ignore other listed profile destinations.
Enable global routing without checking whether any listed destination violates organizational policy.
Grant broad S3 GetObject rights as a remedy for a model profile SCP denial.
Validate the profile’s actual required destination Regions against IAM and SCP policies and select or permit only a compliant profile before deployment.

Explanation

Correct Answer

Cross-Region inference requires permissions for the invoked profile and all necessary destination model resources; preflight the destination map against organization controls instead of waiting for runtime failure.

Incorrect Answers

The cross-Region request can fail when a required destination Region is denied.

A more flexible route can break compliance rather than fix permissions.

S3 permissions cannot override an SCP restriction on Bedrock model invocation.

Pass Your Amazon AWS Certified Generative AI Developer - Professional AIP-C01 Exam Easy!

Question 310

Explanation

Correct Answer

Incorrect Answers

Question 311

Explanation

Correct Answer

Incorrect Answers

Question 312

Explanation

Correct Answer

Incorrect Answers

Question 313

Explanation

Correct Answer

Incorrect Answers

Question 314

Explanation

Correct Answer

Incorrect Answers

Question 315

Explanation

Correct Answer

Incorrect Answers

Question 316

Explanation

Correct Answer

Incorrect Answers

Question 317

Explanation

Correct Answer

Incorrect Answers

Question 318

Explanation

Correct Answer

Incorrect Answers

Question 319

Explanation

Correct Answer

Incorrect Answers

Question 320

Explanation

Correct Answer

Incorrect Answers

AWS Certified Generative AI Developer - Professional AIP-C01 Practice Tests