Amazon AWS Certified Generative AI Developer - Professional AIP-C01 Practice Test 31

Question 300

A generative SQL tool proposes DROP TABLE. Which safety control actually addresses the reported weakness?

Run DROP TABLE whenever the SQL syntax parses and the model says the user requested cleanup.
Give the database connector administrator rights because the model needs flexibility.
Use a prompt instruction 'avoid dangerous SQL' without an executor-side statement policy.
Validate allowed SQL operations and database identity outside the model.

Explanation

Correct Answer

Apply SQL parsing/allowlisting, least-privilege DB identity and explicit destructive-action approval before any model-generated query executes.

Incorrect Answers

Syntax and text intent do not authorize destructive database operations.

Broad DB permissions can bypass safe operation boundaries.

Prompt text cannot enforce allowed database operations at commit time.

Question 301

An agent renders arbitrary HTML returned by a tool; the response contains active script markup from an untrusted source. Select TWO output-boundary measures that prevent the text from becoming executable browser code.

Insert raw HTML from the tool into the page using innerHTML with scripts enabled.
Escape or sanitize untrusted tool text for rendering.
Trust a string marked 'from our service' as safe HTML without escaping.
Rely solely on output content-topic moderation to prevent browser XSS.
Constrain active HTML/script execution under approved content policy.

Explanation

Correct Answer

Escape untrusted tool-returned HTML for the actual renderer to prevent script injection.

Apply restrictive active-content policy/sandbox so unsafe scripts cannot execute even if markup is displayed.

Incorrect Answers

Tool output can contain attacker-supplied executable script.

Service provenance does not prove every returned field is sanitized.

Topic filters are not deterministic HTML/JavaScript encoding and execution controls.

Question 302

A Bedrock Guardrails policy must block disallowed user prompts and disallowed model replies while preserving permitted answers. Which THREE acceptance criteria explicitly cover safe messages and both block decisions?

Enable the filter but allow every blocked user input to proceed to the model as long as it is logged.
Return the original disallowed prompt or model output verbatim whenever a guardrail says the content was blocked.
Configure a safe blocked-prompt message that does not echo the unsafe request.
Configure a safe blocked-response message so disallowed generated content is not delivered to the caller.
Test known allowed and blocked prompts and responses to confirm both correct filtering and no excessive blocking.
Use a guardrail without checking any supported input/model/tier constraints and infer acceptance from a successful deployment command.

Explanation

Correct Answer

Guardrails support configuring an appropriate safe message when input triggers a policy; it should not reproduce disallowed content.

Output blocking must replace or suppress prohibited generated text before the application sends a response.

A release test must include both allowed cases (false-positive control) and disallowed cases on both input and output paths.

Incorrect Answers

Logging an unsafe prompt does not make it acceptable to bypass an input block; the prompt must be handled according to policy.

Echoing blocked input or output defeats the block and may expose content the filter was designed to suppress.

A created policy object does not establish supported feature coverage or correct enforcement behavior for a particular runtime.

Question 303

A guardrail is applied to only the first user turn. Which safety control actually addresses the reported weakness?

Enforce the guardrail consistently for relevant subsequent input/output calls.
Disable input moderation after the initial greeting because the session is already authenticated.
Check each later prompt only for JSON validity instead of content policy.
Presume the model always remembers the first-turn moderation result even after context truncation.

Explanation

Correct Answer

Evaluate relevant user turns and generated responses consistently according to the configured guardrail scope; only protecting turn one leaves later attack paths unfiltered.

Incorrect Answers

Authentication does not imply later content is safe.

Syntax validation does not detect unsafe instructions or output.

A historic response does not enforce future requests in API calls.

Question 304

A protection layer logs blocked sensitive text verbatim. Which enforcement decision best mitigates this failure?

Write full prohibited PII strings verbatim to a broadly accessible log group for every guardrail block.
Mask content only in the customer display but keep raw payloads indefinitely in S3 logs.
Minimize or disable high-risk invocation payload logging while preserving safe metadata.
Remove all trace correlation and keep the full sensitive text so security teams can search it.

Explanation

Correct Answer

Redact or minimize sensitive blocked content in invocation logs while retaining safe timestamps, IDs and policy decisions necessary for audit.

Incorrect Answers

A safety tool then becomes a secondary sensitive-data leak.

UI redaction does not restrict backend retention or access.

Identity-free logs still expose PII and lose useful forensic linkage.

Question 305

A service already uses the standalone Bedrock ApplyGuardrail API to moderate outputs from an external model but receives AccessDenied for its intended guardrail resource. The service does not invoke a Bedrock model. What is the least-privilege remedy?

Grant bedrock:* on all foundation models just to run independent guardrail checks.
Grant only the scoped bedrock:ApplyGuardrail IAM permission for the intended guardrail resource/version and enforce the returned intervention before releasing output.
Call Converse with a dummy user prompt instead of the ApplyGuardrail API.
Allow any caller to apply any unpublished guardrail version without identity checks.

Explanation

Correct Answer

The standalone guardrail operation has its own IAM authorization, independent of model inference. Grant the exact guardrail permission/resource needed; broad InvokeModel rights do not resolve the ApplyGuardrail AccessDenied error.

Incorrect Answers

Excessive model permissions are unnecessary for ApplyGuardrail-only operation.

Dummy generation is a different operation and may incur cost or model-specific restrictions.

Guardrail use still needs scoped authorization and an approved version.

Question 306

A model invocation fails when referencing a guardrail. Which safety control actually addresses the reported weakness?

Validate guardrail ID/version and permissions instead of disabling all safety enforcement.
Remove guardrail configuration entirely and treat the resulting successful inference as equivalent.
Replace a nonexistent guardrail version with a random integer until requests stop failing.
Add wildcard S3 permissions to fix a guardrail ARN validation error.

Explanation

Correct Answer

Check the exact guardrail identifier, available version and role's ApplyGuardrail/inference permissions; keep moderation active while repairing configuration.

Incorrect Answers

A successful unmoderated call is not a safety-gate PASS.

Version and policy must be verified, not guessed.

Object access does not authorize an invalid Guardrails configuration.

Question 307

A new prompt version causes higher toxicity on edge cases. Which enforcement decision best mitigates this failure?

Evaluate labeled adversarial regression suite and block unsafe promotion.
Approve the prompt after one clean benign example because syntax and grammar improved.
Lower model temperature and assume this eliminates all jailbreak responses.
Remove difficult edge cases from the evaluation set so the new prompt reaches target score.

Explanation

Correct Answer

Run labeled toxic/adversarial regression cases against the versioned prompt and block deployment when policy failure rate increases.

Incorrect Answers

Happy-path fluency does not test toxic edge cases.

Sampling adjustments are not policy enforcement.

Discarding failures falsifies comparative safety evidence.

Question 308

A data-extraction agent reveals sensitive fields in generated trace. What should be implemented before letting the model answer or act?

Mask sensitive output only at the final user reply while leaving raw tool traces intact.
Copy all extraction inputs into public debug dashboards to support agent tuning.
Replace customer names but leave unique identifying numbers visible in trace attributes.
Scrub intermediate traces and enforce data minimization in orchestration.

Explanation

Correct Answer

Redact sensitive intermediates in orchestration telemetry, minimize passed fields, and secure storage with appropriate IAM/retention policy.

Incorrect Answers

Intermediate logs can still expose the extracted personal data.

Broad telemetry distribution violates processing-purpose restrictions.

Partial anonymization may still expose high-risk identifiers.

Question 309

A third-party LLM produces a completed text response outside Bedrock, and the application must prevent disallowed output from reaching customers. Native Bedrock model guardrail integration is unavailable. Which output-release control is suitable?

Call the standalone Bedrock ApplyGuardrail API on the external model output and enforce its returned intervention before sending the text to users.
Assume a third-party provider's internal filters are identical to the organization's Bedrock Guardrail policy.
Invoke a Bedrock guardrail by adding its ID as a natural-language line to the external prompt.
Display external model output while awaiting the independent guardrail result.

Explanation

Correct Answer

Independent ApplyGuardrail can moderate externally generated text. Assess the OUTPUT before publication; checking only the input or granting general model-invocation rights would not enforce this output boundary.

Incorrect Answers

Provider moderation does not prove configured organization-specific control.

Prompt content cannot invoke the ApplyGuardrail API.

A delayed decision cannot prevent content already disclosed.

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.

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

Question 300

Explanation

Correct Answer

Incorrect Answers

Question 301

Explanation

Correct Answer

Incorrect Answers

Question 302

Explanation

Correct Answer

Incorrect Answers

Question 303

Explanation

Correct Answer

Incorrect Answers

Question 304

Explanation

Correct Answer

Incorrect Answers

Question 305

Explanation

Correct Answer

Incorrect Answers

Question 306

Explanation

Correct Answer

Incorrect Answers

Question 307

Explanation

Correct Answer

Incorrect Answers

Question 308

Explanation

Correct Answer

Incorrect Answers

Question 309

Explanation

Correct Answer

Incorrect Answers

Question 310

Explanation

Correct Answer

Incorrect Answers

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