Amazon AWS Certified Generative AI Developer - Professional AIP-C01 Practice Test 30

Question 290

An enterprise assistant receives an instruction to ignore its system policy and reveal another employee's account. Which protection combines an appropriate Bedrock prompt-attack filter with the essential independent access boundary?

Rely on a low model temperature as proof that the instruction-override attack will not succeed.
Configure supported Bedrock Guardrails prompt-attack detection and independently enforce caller authorization on protected data/tools.
Tell the attacker which sensitive records exist so the model can explain why it refuses.
Grant the model an administrator tool role but ask politely that it not use the role.

Explanation

Correct Answer

Guardrails can help detect prompt attacks, while the application or tool still must validate identity and record-level authorization regardless of model output.

Incorrect Answers

Sampling settings do not detect jailbreaks or enforce the caller's resource permissions.

Exposing the existence or details of restricted records can itself leak sensitive information.

A broad role remains broad even with benign system instructions; a prompt cannot substitute for least-privilege access controls.

Question 291

A model emits a dangerous procedural response. Which safety control actually addresses the reported weakness?

Mark the request safe because its user input contains no prohibited words.
Evaluate the generated output with a guardrail rather than checking inputs only.
Apply the guardrail only when the user explicitly asks for moderation.
Return the unsafe output and attach a warning badge afterward.

Explanation

Correct Answer

Evaluate generated output as OUTPUT with the configured guardrail and intercept dangerous content before serving it, not only input prompts.

Incorrect Answers

Models can independently generate policy-violating text.

Output safety is a consistent application control, not an opt-in courtesy.

Post-disclosure labeling does not prevent harmful content exposure.

Question 292

A single-turn policy-document question-answering workflow supplies an authoritative source passage and a user query, but its generated answer invents a clause missing from that source. This is not a conversational chatbot. What check is appropriate for this supported use case?

Run only a semantic similarity score between the question and answer, without comparing the claimed clause with source text.
Inspect request token counts as a proxy for whether generated policy clauses are true.
Use standard profanity filters and assume they detect unsupported factual statements.
Apply a supported contextual grounding check to the source, query and generated response, and separately verify cited clauses in the authoritative source.

Explanation

Correct Answer

For supported source/query/answer cases, contextual grounding can flag ungrounded claims; verifying cited document evidence reduces false confidence. AWS does not support this specific feature for conversational QA/chatbot cases.

Incorrect Answers

A reply may be relevant to the question while introducing an unsupported clause. Relevance is not factual grounding.

Token count tracks usage, not whether claims are entailed by the reference passage.

Harmful-content categories do not prove that factual statements are supported by an approved policy.

Question 293

A team interprets a high retrieval-similarity score as proof that the model both answered the user's question AND did not invent facts. Which TWO distinct evaluations test those two claims independently?

Replace both checks with a single embedding similarity number and label it a factual-accuracy score.
Measure only inference latency and token count because a fast answer is more likely to be true.
Remove low-scoring documents until every returned passage appears semantically similar, without checking generated text.
Evaluate whether the answer addresses the actual user query (relevance).
Evaluate whether each factual assertion is supported by the supplied authoritative source (grounding).

Explanation

Correct Answer

An answer can be supported by a document yet ignore what the user asked. Relevance tests query-to-response alignment.

An answer can address the question but invent a source clause. Grounding tests response assertions against the reference evidence.

Incorrect Answers

Embedding similarity can identify related content, not establish that the answer is factually entailed by it or responsive to the question.

Operational metrics do not determine relevance or source-supported truth.

Retriever rank and factual verification of model-generated claims are separate stages; deleting candidates cannot replace either review.

Question 294

A conversational assistant uses contextual grounding check for unsupported chatbot mode. Which safety control actually addresses the reported weakness?

Enable contextual grounding on every multi-turn chat and claim universal factual coverage.
Use a topic-denial score as a substitute for source-grounded claim checking.
Verify Bedrock grounding-check supported use cases before claiming coverage.
Treat a high overall relevance score as proof every claim is supported by a source.

Explanation

Correct Answer

AWS documents contextual grounding for source-based summarization/paraphrasing/QA with supplied source/query, and excludes general conversational QA/chatbot use cases in this guide; validate exact supported feature scope.

Incorrect Answers

The documented scope has restrictions and may not support that chatbot mode.

Topic filtering and factual entailment are different policies.

Relevance and grounding are separate evaluations and may diverge.

Question 295

A guardrail blocks content but the application loses user context. Which enforcement decision best mitigates this failure?

Return configured block message and safe escalation while preserving authorized session state.
Reset every user's session and expose a fresh anonymous chat after any blocked request.
Display the original blocked harmful output alongside the safety warning for context.
Treat a guardrail refusal as authorization to skip outstanding human-review requirements.

Explanation

Correct Answer

Return the configured guardrail intervention message and safe next action without erasing authenticated, authorized session context.

Incorrect Answers

Losing verified state is unnecessary and can damage workflow continuity.

The blocked content must not be returned.

A blocked response has no authority to clear business controls.

Question 296

A model produces a JSON object with an unauthorized command. What should be implemented before letting the model answer or act?

Execute every syntactically valid JSON command because the text passed moderation.
Have the model label dangerous commands as safe and rely on that output field.
Use profanity filtering as the only control before forwarding commands to a payment API.
Apply schema and policy allowlists in post-processing independent of content filtering.

Explanation

Correct Answer

A content guardrail cannot authorize executable commands; enforce typed action allowlists and business policy after parsing the model's JSON.

Incorrect Answers

Content safety does not imply permission for database operations.

Model-generated policy labels are not a trusted action authorization.

Text moderation cannot constrain business operation type or account scope.

Question 297

A malicious PDF contains instructions to exfiltrate a database. Which safety control actually addresses the reported weakness?

Promote instructions found in PDFs to system-message priority after retrieval.
Treat document instructions as untrusted evidence and prohibit tool permission escalation.
Run the requested database export because the document claims it is an admin playbook.
Ask the model to decide if permission checks are necessary for this particular document.

Explanation

Correct Answer

PDF content is untrusted data: retrieve for evidence but never let it redefine tool permissions or authorize database exfiltration.

Incorrect Answers

An attacker-controlled document can then hijack trusted policy.

Document text cannot authenticate the user or grant data-export rights.

Authorization must be enforced outside model discretion.

Question 298

A jailbreak attempts role-play to bypass approved safety controls. Which enforcement decision best mitigates this failure?

Disable safety controls when a test request says the interaction is fictional.
Approve the model because it refused one fixed test phrase.
Run adversarial evaluation and enforce external moderation/permission checks.
Let the user declare themselves an internal tester and acquire elevated tool privileges automatically.

Explanation

Correct Answer

Use a labeled adversarial suite including roleplay jailbreaks and enforce independent guardrail/tool permission boundaries under those prompts.

Incorrect Answers

Roleplay framing is itself a common jailbreak tactic.

One refusal does not establish resistance to broader variations.

A chat assertion does not grant tested roles or execution rights.

Question 299

A model returns an unsupported medical treatment classification. What should be implemented before letting the model answer or act?

Return whichever treatment category sounds most likely based on answer fluency.
Convert model confidence into a definitive medical decision without calibration or human checks.
Ignore uncertainty because a JSON schema requires a non-null treatment category.
Escalate safety-critical decisions instead of fabricating a deterministic safe label.

Explanation

Correct Answer

For safety-critical medical classifications lacking reliable validated evidence, abstain/escalate rather than invent a diagnostic treatment label.

Incorrect Answers

A persuasive answer is not a clinical validation.

Uncalibrated confidence is unsuitable for safety-critical action.

A schema constraint must not force fabrication of unsupported medical guidance.

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.

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

Question 290

Explanation

Correct Answer

Incorrect Answers

Question 291

Explanation

Correct Answer

Incorrect Answers

Question 292

Explanation

Correct Answer

Incorrect Answers

Question 293

Explanation

Correct Answer

Incorrect Answers

Question 294

Explanation

Correct Answer

Incorrect Answers

Question 295

Explanation

Correct Answer

Incorrect Answers

Question 296

Explanation

Correct Answer

Incorrect Answers

Question 297

Explanation

Correct Answer

Incorrect Answers

Question 298

Explanation

Correct Answer

Incorrect Answers

Question 299

Explanation

Correct Answer

Incorrect Answers

Question 300

Explanation

Correct Answer

Incorrect Answers

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