Topic 09 Practice Test 1 covers Guardrails, Prompt-Injection Defense, Output Safety, Grounding, and Hallucination Controls in Domain 3 (Task 3.1). For broader exam preparation, review the AIP-C01 Exam Dumps. Every option includes focused technical reasoning tied to the tested decision.
Question 1
A public support assistant must reject requests containing severe hate or violence before the prompt reaches the foundation model. The team wants a managed Bedrock control with category strengths rather than custom code. Which implementation best meets the requirement? Choose ONE.
- Amazon Bedrock Guardrails denied topics
- Step Functions and Lambda custom moderation workflow
- Amazon Bedrock Guardrails input content filters
- Amazon Bedrock Guardrails word filters
Correct Answer: C
Correct Answer
Answer C is correct because Amazon Bedrock Guardrails input content filters matches both the mechanism and placement required here. Bedrock Guardrails content filters classify configured harmful categories on incoming content before inference. The scenario calls for managed category-based harmful-input filtering; the filter strength can be tuned by policy instead of embedding moderation logic in every application path.
Incorrect Answers
Answer A is incorrect because Amazon Bedrock Guardrails denied topics belongs in designs that need blocking semantically defined subjects that the application is not allowed to discuss, such as investment advice for a retail-banking bot. The present requirement is narrower: managed category-based harmful-input filtering.
Answer B is incorrect because Step Functions and Lambda custom moderation workflow is intended for orchestrating organization-specific validation, enrichment, approvals, and branching when a single managed filter is insufficient. Since it adds workflow logic rather than serving as a simple managed content-category filter, it would shift attention to a different control objective and leave the stated requirement unsatisfied.
Answer D is incorrect because Amazon Bedrock Guardrails word filters serves blocking exact custom words or phrases, including organization-specific terms that must never appear in an interaction. Since it does not reason about a broad topic semantically, it would not enforce the exact boundary described and therefore is not the best answer.
Question 2
A consumer assistant occasionally generates abusive language even from benign prompts. The response must be checked against managed harmful-content categories before users can see it. Which control should be applied? Choose ONE.
- Bedrock Guardrails contextual grounding check
- JSON Schema constrained output validation
- Specialized toxicity and content-moderation evaluation
- Amazon Bedrock Guardrails output content filters
Correct Answer: D
Correct Answer
Answer D is correct because The best answer is Amazon Bedrock Guardrails output content filters. That directly implements managed harmful-output filtering, and grounding or schema validation addresses different failure modes and would not directly block toxic language. The distinction matters because a control aimed at a different layer could leave the specified exposure unchanged.
Incorrect Answers
Answer A is incorrect because Bedrock Guardrails contextual grounding check can strengthen a system through scoring whether a response is supported by supplied grounding source content and filtering responses below the configured grounding threshold. Yet the required outcome here is managed harmful-output filtering.
Answer B is incorrect because JSON Schema constrained output validation solves problems involving requiring generated structured output to conform to expected fields, types, and allowed shapes before downstream automation accepts it. Its limitation is that schema validity cannot prove that a semantically plausible value is factually true, so the application would still be missing the mechanism the question asks for.
Answer C is incorrect because Specialized toxicity and content-moderation evaluation addresses measuring or classifying generated outputs for toxic or unsafe language so the application can block, score, or route them. Since it does not enforce relational database semantics, selecting it would protect a different boundary and would not close the specific exposure described.
Question 3
An employee assistant answers questions from a controlled policy library. The team wants the model to retrieve current passages and use them as context instead of relying on memorized training data. Which AWS feature best supports this design? Choose ONE.
- Authoritative runtime data retrieval
- Amazon Bedrock Knowledge Bases retrieval grounding
- Bedrock Guardrails contextual grounding check
- Evidence-backed source citation
Correct Answer: B
Correct Answer
Answer B is correct because Amazon Bedrock Knowledge Bases retrieval grounding is appropriate because Amazon Bedrock Knowledge Bases retrieves relevant source material and supplies it to the model as grounding context for RAG. the retrieved evidence can then be combined with additional grounding checks when stricter verification is needed.
Incorrect Answers
Answer A is incorrect because Authoritative runtime data retrieval is appropriate for querying the current system of record at request time instead of asking the model to recall volatile facts from training. Since it does not by itself filter harmful language or prompt attacks, it belongs elsewhere in the design and does not meet the scenario’s primary safety objective.
Answer C is incorrect because Bedrock Guardrails contextual grounding check supports scoring whether a response is supported by supplied grounding source content and filtering responses below the configured grounding threshold, which is useful in the right context. The deciding requirement here is managed retrieval grounding from enterprise documents.
Answer D is incorrect because Evidence-backed source citation belongs in designs that need returning identifiers or references for the retrieved material supporting an answer so users or downstream checks can verify important claims. The present requirement is narrower: managed retrieval grounding from enterprise documents.
Question 4
A public GenAI API must use separate safeguards before the model, at the model boundary, after generation, and at the outbound API edge. Which architecture best reflects defense in depth? Choose ONE.
- ApplyGuardrail on both input and output
- Step Functions safety orchestration
- Independent safety classifier layer
- Comprehend → Guardrails → Lambda → API Gateway layered safety path
Correct Answer: D
Correct Answer
Answer D is correct because This requirement points to Comprehend → Guardrails → Lambda → API Gateway layered safety path: a Comprehend pre-filter, Bedrock guardrail, Lambda post-validator, and API-edge check create distinct layers with different failure coverage. The necessary capability is multiple independent controls across the request-response path. no single classifier becomes the only safety boundary.
Incorrect Answers
Answer A is incorrect because ApplyGuardrail on both input and output can be valuable when implementing running the configured guardrail before inference and again on generated content so unsafe prompts and unsafe responses are independently evaluated. The current scenario is determined by multiple independent controls across the request-response path.
Answer B is incorrect because Step Functions safety orchestration can strengthen a system through coordinating multiple safety checks, retries, approval branches, and fail-closed handling as an explicit workflow. Because it orchestrates controls rather than replacing the controls it invokes, this choice would address a neighboring risk and leave the central requirement unmet.
Answer C is incorrect because Independent safety classifier layer is a reasonable control for using a classifier separate from the primary generator so a single model is not solely responsible for judging its own output. However, this workload specifically calls for multiple independent controls across the request-response path.
Question 5
A user tells a banking assistant to ignore its developer instructions, adopt a different role, and perform a task outside the assistant’s intended workflow. The application needs a managed Bedrock control designed for this override attempt. Which feature should be enabled? Choose ONE.
- Amazon Bedrock Guardrails input content filters
- Amazon Bedrock Guardrails prompt attack filter
- Deterministic request schema validation
- Amazon Bedrock Guardrails denied topics
Correct Answer: B
Correct Answer
Answer B is correct because Amazon Bedrock Guardrails prompt attack filter fits the evidence in the scenario because the Bedrock Guardrails prompt attack filter is designed to identify instructions that try to override developer guidance or hijack the intended task. this targets adversarial instruction override more directly than ordinary harmful-content or request-shape checks.
Incorrect Answers
Answer A is incorrect because Amazon Bedrock Guardrails input content filters can help with classifying and blocking harmful categories such as hate, violence, sexual content, insults, misconduct, or prompt attacks before a supported model processes the request, but the question hinges on managed detection of prompt-injection attempts. Its limitation here is that it is not a custom multi-step business escalation workflow.
Answer C is incorrect because Deterministic request schema validation is appropriate for requiring the incoming payload to match allowed fields, types, sizes, and enum values before any prompt is assembled. Since it does not detect adversarial natural-language intent by itself, it belongs elsewhere in the design and does not meet the scenario’s primary safety objective.
Answer D is incorrect because The value of Amazon Bedrock Guardrails denied topics is blocking semantically defined subjects that the application is not allowed to discuss, such as investment advice for a retail-banking bot. Since it is not intended to match one literal word or entity name, the design would still need a different control to satisfy the question.
Question 6
A retail-bank chatbot may answer account questions but must not discuss cryptocurrency or investment recommendations, even when the user phrases the request indirectly. Which safety control is the most appropriate? Choose ONE.
- Amazon Bedrock Guardrails input content filters
- Amazon Bedrock Guardrails denied topics
- Amazon Bedrock Guardrails word filters
- Deterministic request schema validation
Correct Answer: B
Correct Answer
Answer B is correct because Use Amazon Bedrock Guardrails denied topics. denied topics evaluate whether the request belongs to a prohibited theme rather than relying on exact keywords. The decisive phrase in the requirement is semantic blocking of a prohibited subject, and this continues to work when wording changes but the disallowed subject remains the same.
Incorrect Answers
Answer A is incorrect because Amazon Bedrock Guardrails input content filters is best suited to classifying and blocking harmful categories such as hate, violence, sexual content, insults, misconduct, or prompt attacks before a supported model processes the request. The deciding need here is semantic blocking of a prohibited subject.
Answer C is incorrect because Amazon Bedrock Guardrails word filters primarily supports blocking exact custom words or phrases, including organization-specific terms that must never appear in an interaction. Its shortcoming is that it does not reason about a broad topic semantically, so choosing it would leave the requested safety or accuracy mechanism incomplete.
Answer D is incorrect because Deterministic request schema validation can be valuable when implementing requiring the incoming payload to match allowed fields, types, sizes, and enum values before any prompt is assembled. Because it does not detect adversarial natural-language intent by itself, it cannot provide the required behavior without an additional, more directly targeted safeguard.
Question 7
A self-hosted model produces text outside Bedrock, but the organization wants to evaluate each completed response with its existing Bedrock guardrail before returning it. What should the application call? Choose ONE.
- API Gateway response filtering layer
- Amazon Bedrock Guardrails output content filters
- Lambda post-processing safety validator
- ApplyGuardrail on generated output
Correct Answer: D
Correct Answer
Answer D is correct because ApplyGuardrail on generated output provides the required mechanism because ApplyGuardrail can inspect already-generated content without performing the model invocation, making it suitable around an external model. The operational objective is independent guardrail evaluation of generated text. the application can enforce the returned guardrail action before the client receives the response.
Incorrect Answers
Answer A is incorrect because API Gateway response filtering layer can help with enforcing a final API-edge response policy or transformation after application processing and before data leaves the service boundary, but the question hinges on independent guardrail evaluation of generated text. Its limitation here is that it is not a model-aware hallucination detector by itself.
Answer B is incorrect because Amazon Bedrock Guardrails output content filters is relevant to evaluating generated text against harmful-content categories before the response is delivered. Because it does not prove that every factual claim is supported by a reference source, the option is technically useful but positioned for a different purpose than the control being requested.
Answer C is incorrect because Lambda post-processing safety validator has value for running custom deterministic checks on generated output before it reaches the caller when managed policies are not sufficient. Because it requires application logic and does not automatically provide all Bedrock Guardrail policy types, the choice does not align with the point where the specified risk must be controlled.
Question 8
A summarization service has the original document, the user’s request, and the generated summary. It must block a summary that introduces facts absent from the source. Which Bedrock Guardrails feature is designed for this? Choose ONE.
- Bedrock Guardrails contextual grounding check
- Amazon Bedrock Guardrails output content filters
- Bedrock Guardrails contextual relevance check
- Semantic-similarity evidence verification
Correct Answer: A
Correct Answer
Answer A is correct because The scenario is best handled by Bedrock Guardrails contextual grounding check. contextual grounding compares the response with the grounding source and can filter output below a configured grounding threshold. Its relevance comes from the need for grounding-score verification against supplied source content; this directly targets unsupported additions rather than general harmful-content categories.
Incorrect Answers
Answer B is incorrect because Amazon Bedrock Guardrails output content filters is useful when evaluating generated text against harmful-content categories before the response is delivered. Because it does not prove that every factual claim is supported by a reference source, it would leave the deciding safety requirement unresolved even though it could still appear elsewhere in a layered design.
Answer C is incorrect because Bedrock Guardrails contextual relevance check helps with scoring whether a generated response actually addresses the supplied user query relative to the provided context. Because relevance alone does not establish that every statement is factually grounded, the option would not produce the requested control outcome, though it can remain part of a broader defense strategy.
Answer D is incorrect because Semantic-similarity evidence verification primarily supports comparing generated claims or answer content with authoritative retrieved evidence and rejecting material mismatches. Its shortcoming is that similarity is a verification signal and should not be treated as a replacement for source governance, so choosing it would leave the requested safety or accuracy mechanism incomplete.
Question 9
A team uses ApplyGuardrail only on user prompts. Testing shows benign prompts can still produce unsafe responses. What change most directly strengthens the existing design? Choose ONE.
- ApplyGuardrail on both input and output
- Amazon Bedrock Guardrails output content filters
- Comprehend → Guardrails → Lambda → API Gateway layered safety path
- Fail-closed safety handling
Correct Answer: A
Correct Answer
Answer A is correct because ApplyGuardrail on both input and output is the strongest fit because running ApplyGuardrail before inference and again on the generated content covers unsafe inputs and unsafe outputs separately. The deciding requirement is independent safety checks on both sides of inference. the same policy can therefore protect both directions even when the model itself behaves unexpectedly.
Incorrect Answers
Answer B is incorrect because Amazon Bedrock Guardrails output content filters fits situations involving evaluating generated text against harmful-content categories before the response is delivered. Its limitation here is that it does not prove that every factual claim is supported by a reference source, making it complementary rather than sufficient for the requested control.
Answer C is incorrect because Use Comprehend → Guardrails → Lambda → API Gateway layered safety path when the goal is using distinct pre-processing, managed guardrails, custom post-processing, and API-edge controls so one failure does not become the only barrier. That is not the decisive goal in this scenario, which is independent safety checks on both sides of inference.
Answer D is incorrect because Fail-closed safety handling is relevant to withholding a response or action when a required safety control errors, times out, or returns an indeterminate result. Because it governs failure behavior rather than detecting a specific unsafe category, the option is technically useful but positioned for a different purpose than the control being requested.
Question 10
A Bedrock prompt-attack filter is already enabled, but the security team also requires an independent second detector trained on the organization’s role-play and ‘developer mode’ jailbreak corpus. Which added capability best meets that requirement? Choose ONE.
- Safety telemetry with automated alerting
- Jailbreak-focused adversarial detector
- Amazon Bedrock Guardrails word filters
- Bedrock Guardrails contextual grounding check
Correct Answer: B
Correct Answer
Answer B is correct because The required control is Jailbreak-focused adversarial detector. Here, a jailbreak-focused detector can score organization-specific coercion and role-manipulation patterns as a second signal before inference, which matches the need for an independent jailbreak-specific detection signal. This is more precise than selecting a neighboring safeguard that protects a different stage or failure mode.
Incorrect Answers
Answer A is incorrect because Safety telemetry with automated alerting helps with recording guardrail interventions, classifier scores, and blocked attack patterns so operators can detect changing abuse and tune controls. Because observability detects trends but should not be the only runtime blocking control, the option would not produce the requested control outcome, though it can remain part of a broader defense strategy.
Answer C is incorrect because The normal role of Amazon Bedrock Guardrails word filters is blocking exact custom words or phrases, including organization-specific terms that must never appear in an interaction. Because it does not reason about a broad topic semantically, this option solves an adjacent problem rather than the one that determines the answer.
Answer D is incorrect because Bedrock Guardrails contextual grounding check is designed around scoring whether a response is supported by supplied grounding source content and filtering responses below the configured grounding threshold. The workload here is asking for an independent jailbreak-specific detection signal.
Question 11
An enterprise assistant must block the exact internal codename ORCHID-17 whenever it appears in user input. The requirement is literal matching, not semantic topic detection. What should be configured? Choose ONE.
- Amazon Bedrock Guardrails input content filters
- Amazon Bedrock Guardrails denied topics
- Amazon Bedrock Guardrails word filters
- Guardrails sensitive-information custom regex filter
Correct Answer: C
Correct Answer
Answer C is correct because Amazon Bedrock Guardrails word filters directly satisfies this scenario: word filters are designed for custom exact words and phrases that must be blocked in interactions. The application specifically needs exact blocking of a specific phrase. using a topic classifier would be unnecessarily broad for a single literal codename.
Incorrect Answers
Answer A is incorrect because Amazon Bedrock Guardrails input content filters supports classifying and blocking harmful categories such as hate, violence, sexual content, insults, misconduct, or prompt attacks before a supported model processes the request. That capability does not satisfy exact blocking of a specific phrase.
Answer B is incorrect because Amazon Bedrock Guardrails denied topics would make sense for blocking semantically defined subjects that the application is not allowed to discuss, such as investment advice for a retail-banking bot. In the stated scenario, the determining constraint is exact blocking of a specific phrase.
Answer D is incorrect because Guardrails sensitive-information custom regex filter fits situations involving detecting organization-specific structured identifiers that follow a known pattern and then blocking or masking them. Its limitation here is that it is not the right control for a semantically defined prohibited subject, making it complementary rather than sufficient for the requested control.
Question 12
A release pipeline needs a quantitative safety check that scores batches of candidate responses for toxicity before a new prompt version can be promoted. Which approach best fits? Choose ONE.
- Human review for high-impact generated output
- Specialized toxicity and content-moderation evaluation
- Deterministic output allowlist validation
- Amazon Bedrock Guardrails output content filters
Correct Answer: B
Correct Answer
Answer B is correct because Choose Specialized toxicity and content-moderation evaluation because a specialized toxicity or moderation evaluation produces measurable safety evidence over representative generated outputs. In this workload, the critical constraint is batch toxicity evaluation for release evidence. a runtime blocking guardrail is still useful, but it does not by itself provide the requested comparative release metric.
Incorrect Answers
Answer A is incorrect because Human review for high-impact generated output is intended for placing a qualified reviewer before release when automated safety checks are insufficient for consequential or ambiguous content. Since it introduces latency and is not the first-line automated filter for routine traffic, it would shift attention to a different control objective and leave the stated requirement unsatisfied.
Answer C is incorrect because The normal role of Deterministic output allowlist validation is accepting only explicitly permitted values, actions, or categories from model output before downstream execution. Because it is unsuitable for broad free-form safety classification on its own, this option solves an adjacent problem rather than the one that determines the answer.
Answer D is incorrect because Amazon Bedrock Guardrails output content filters serves evaluating generated text against harmful-content categories before the response is delivered. Since it does not prove that every factual claim is supported by a reference source, it would not enforce the exact boundary described and therefore is not the best answer.
Question 13
A model returns an answer that is fully supported by the supplied document but discusses the wrong section and does not answer the user’s question. Which contextual check should detect this problem? Choose ONE.
- Bedrock Guardrails contextual grounding check
- JSON Schema constrained output validation
- Confidence or uncertainty scoring with a decision threshold
- Bedrock Guardrails contextual relevance check
Correct Answer: D
Correct Answer
Answer D is correct because Bedrock Guardrails contextual relevance check matches both the mechanism and placement required here. the contextual relevance check scores whether the generated content addresses the query in the context of the supplied evidence. The scenario calls for response relevance to the user query; a statement can be grounded yet still fail because it answers the wrong question.
Incorrect Answers
Answer A is incorrect because Bedrock Guardrails contextual grounding check solves problems involving scoring whether a response is supported by supplied grounding source content and filtering responses below the configured grounding threshold. Here the exact requirement is response relevance to the user query.
Answer B is incorrect because JSON Schema constrained output validation would make sense for requiring generated structured output to conform to expected fields, types, and allowed shapes before downstream automation accepts it. In the stated scenario, the determining constraint is response relevance to the user query.
Answer C is incorrect because Confidence or uncertainty scoring with a decision threshold addresses estimating answer reliability and using a calibrated threshold to accept, retry, abstain, or escalate uncertain responses. Since an uncalibrated score can create false assurance, selecting it would protect a different boundary and would not close the specific exposure described.
Question 14
A high-risk assistant needs to run input classification, guardrail evaluation, a custom business-policy check, optional human approval, and then output verification with retries. Which AWS service is best suited to coordinate this sequence? Choose ONE.
- Step Functions safety orchestration
- Comprehend → Guardrails → Lambda → API Gateway layered safety path
- API Gateway response filtering layer
- Step Functions and Lambda custom moderation workflow
Correct Answer: A
Correct Answer
Answer A is correct because The best answer is Step Functions safety orchestration. Step Functions can model ordered checks, branching, retries, timeouts, and escalation paths so the complete safety process is visible and controlled. That directly implements explicit orchestration of several safety stages, and the workflow invokes specialized controls instead of overloading one function with every responsibility.
Incorrect Answers
Answer B is incorrect because Comprehend → Guardrails → Lambda → API Gateway layered safety path belongs in designs that need using distinct pre-processing, managed guardrails, custom post-processing, and API-edge controls so one failure does not become the only barrier. The present requirement is narrower: explicit orchestration of several safety stages.
Answer C is incorrect because API Gateway response filtering layer serves enforcing a final API-edge response policy or transformation after application processing and before data leaves the service boundary. Since it is not a model-aware hallucination detector by itself, it would not enforce the exact boundary described and therefore is not the best answer.
Answer D is incorrect because Step Functions and Lambda custom moderation workflow supports orchestrating organization-specific validation, enrichment, approvals, and branching when a single managed filter is insufficient, which is useful in the right context. The deciding requirement here is explicit orchestration of several safety stages.
Question 15
An application accepts Markdown that can contain hidden HTML comments and control-like fragments. Security wants those fragments neutralized before the text is passed to safety classifiers or the model. Which step is most appropriate? Choose ONE.
- Deterministic request schema validation
- Input normalization before safety evaluation
- Untrusted-input sanitization
- Amazon Bedrock Guardrails prompt attack filter
Correct Answer: C
Correct Answer
Answer C is correct because Untrusted-input sanitization is appropriate because input sanitization removes or neutralizes syntax that should be treated as data rather than executable or instruction-like control text. The question is not asking for generic hardening; it asks for sanitize untrusted markup before safety evaluation. semantic attack detection should still run afterward because sanitization alone cannot infer intent.
Incorrect Answers
Answer A is incorrect because Deterministic request schema validation solves problems involving requiring the incoming payload to match allowed fields, types, sizes, and enum values before any prompt is assembled. Its limitation is that it does not detect adversarial natural-language intent by itself, so the application would still be missing the mechanism the question asks for.
Answer B is incorrect because Input normalization before safety evaluation is a reasonable control for canonicalizing encodings, whitespace, and equivalent representations before classifiers and filters inspect the request. The mismatch is that normalization alone does not decide whether the normalized request is safe; therefore the required safety decision would still depend on another component.
Answer D is incorrect because Amazon Bedrock Guardrails prompt attack filter can strengthen a system through detecting prompt injection and jailbreak attempts intended to override developer instructions, bypass safeguards, or expose protected instructions. Yet the required outcome here is sanitize untrusted markup before safety evaluation.
Question 16
A company uses a third-party model endpoint but wants to run the same Bedrock safety policy on each user message before sending it to that endpoint. Which approach provides the cleanest separation? Choose ONE.
- Amazon Bedrock Guardrails input content filters
- Bedrock Guardrails guardContent scoping
- Step Functions and Lambda custom moderation workflow
- ApplyGuardrail before model invocation
Correct Answer: D
Correct Answer
Answer D is correct because This requirement points to ApplyGuardrail before model invocation: ApplyGuardrail can assess text without invoking a Bedrock foundation model, so it can wrap an external inference path. The necessary capability is independent pre-inference guardrail evaluation. the guardrail decision can be enforced before any message is forwarded to the third-party provider.
Incorrect Answers
Answer A is incorrect because The value of Amazon Bedrock Guardrails input content filters is classifying and blocking harmful categories such as hate, violence, sexual content, insults, misconduct, or prompt attacks before a supported model processes the request. Since it is not a custom multi-step business escalation workflow, the design would still need a different control to satisfy the question.
Answer B is incorrect because Bedrock Guardrails guardContent scoping supports marking the specific message blocks that should be evaluated by configured guardrail policies when using supported Bedrock APIs, which is useful in the right context. Since it is a scoping mechanism rather than an independent classifier, it addresses a separate concern and is not the control that closes this gap.
Answer C is incorrect because Step Functions and Lambda custom moderation workflow is appropriate for orchestrating organization-specific validation, enrichment, approvals, and branching when a single managed filter is insufficient. Since it adds workflow logic rather than serving as a simple managed content-category filter, it belongs elsewhere in the design and does not meet the scenario’s primary safety objective.
Question 17
An analyst asks natural-language questions whose answers must come strictly from an approved relational database. Leadership prefers a deterministic database result over a free-form model guess. Which pattern most directly reduces invented answers? Choose ONE.
- JSON Schema constrained output validation
- Amazon Bedrock Knowledge Bases retrieval grounding
- Constrained text-to-SQL with deterministic execution
- Bedrock Guardrails contextual grounding check
Correct Answer: C
Correct Answer
Answer C is correct because Constrained text-to-SQL with deterministic execution fits the evidence in the scenario because a constrained text-to-SQL path can translate intent into a validated query and return records calculated by the database. this is stronger than unconstrained generation when the authoritative answer already exists in structured tables.
Incorrect Answers
Answer A is incorrect because JSON Schema constrained output validation can be valuable when implementing requiring generated structured output to conform to expected fields, types, and allowed shapes before downstream automation accepts it. Because schema validity cannot prove that a semantically plausible value is factually true, it cannot provide the required behavior without an additional, more directly targeted safeguard.
Answer B is incorrect because Amazon Bedrock Knowledge Bases retrieval grounding is best suited to retrieving relevant enterprise source material and supplying it as context so the model answers from managed evidence. The deciding need here is deterministic answers from validated database execution.
Answer D is incorrect because Bedrock Guardrails contextual grounding check is a reasonable control for scoring whether a response is supported by supplied grounding source content and filtering responses below the configured grounding threshold. However, this workload specifically calls for deterministic answers from validated database execution.
Question 18
A grounded-QA application is allowing too many weakly supported responses. Offline evaluation shows that valid answers usually score above 0.82, while risky answers cluster lower. What configuration change should be considered? Choose ONE.
- Bedrock Guardrails contextual grounding check
- Amazon Bedrock Knowledge Bases retrieval grounding
- Tuned contextual-grounding threshold
- Confidence or uncertainty scoring with a decision threshold
Correct Answer: C
Correct Answer
Answer C is correct because Use Tuned contextual-grounding threshold. raising the contextual-grounding threshold after validation can block more low-scoring responses when the current threshold is too permissive. The decisive phrase in the requirement is calibrated tightening of the grounding acceptance threshold, and the change should be based on measured false-accept and false-reject trade-offs rather than an arbitrary maximum.
Incorrect Answers
Answer A is incorrect because Bedrock Guardrails contextual grounding check has value for scoring whether a response is supported by supplied grounding source content and filtering responses below the configured grounding threshold. In contrast, this scenario depends on calibrated tightening of the grounding acceptance threshold.
Answer B is incorrect because The value of Amazon Bedrock Knowledge Bases retrieval grounding is retrieving relevant enterprise source material and supplying it as context so the model answers from managed evidence. That does not match the key requirement, calibrated tightening of the grounding acceptance threshold.
Answer D is incorrect because Confidence or uncertainty scoring with a decision threshold can help with estimating answer reliability and using a calibrated threshold to accept, retry, abstain, or escalate uncertain responses, but the question hinges on calibrated tightening of the grounding acceptance threshold. Its limitation here is that an uncalibrated score can create false assurance.
Question 19
Security does not want the primary language model to be the only component deciding whether its own response is safe. Which design adds an independent judgment layer? Choose ONE.
- Independent safety classifier layer
- Confidence or uncertainty scoring with a decision threshold
- Human review for high-impact generated output
- Amazon Bedrock Guardrails output content filters
Correct Answer: A
Correct Answer
Answer A is correct because Independent safety classifier layer provides the required mechanism because an independent classifier provides a second judgment path that is operationally distinct from the generator, reducing reliance on self-evaluation. The operational objective is a separate safety decision signal. its threshold and failure behavior must be validated like any other production safety control.
Incorrect Answers
Answer B is incorrect because Confidence or uncertainty scoring with a decision threshold primarily supports estimating answer reliability and using a calibrated threshold to accept, retry, abstain, or escalate uncertain responses. Its shortcoming is that an uncalibrated score can create false assurance, so choosing it would leave the requested safety or accuracy mechanism incomplete.
Answer C is incorrect because Human review for high-impact generated output is best suited to placing a qualified reviewer before release when automated safety checks are insufficient for consequential or ambiguous content. Its gap is that it introduces latency and is not the first-line automated filter for routine traffic; consequently, it cannot replace the control that directly addresses this condition.
Answer D is incorrect because Amazon Bedrock Guardrails output content filters is useful when evaluating generated text against harmful-content categories before the response is delivered. This case instead requires a separate safety decision signal. Because it does not prove that every factual claim is supported by a reference source, it would leave the deciding safety requirement unresolved even though it could still appear elsewhere in a layered design.
Question 20
A team wants every build to be tested automatically with known prompt-injection, jailbreak, obfuscation, and policy-evasion cases before deployment. Which control best meets the requirement? Choose ONE.
- Automated adversarial testing workflow
- Red-team regression suite in CI/CD
- Amazon Bedrock Guardrails prompt attack filter
- Production canary attack probes and monitoring
Correct Answer: A
Correct Answer
Answer A is correct because The scenario is best handled by Automated adversarial testing workflow. an adversarial test workflow exercises many attack families and records whether controls still resist them, exposing regressions before production. Its relevance comes from the need for automated adversarial security testing before release; runtime filters remain necessary because testing cannot enumerate every future attack.
Incorrect Answers
Answer B is incorrect because Use Red-team regression suite in CI/CD when the goal is rerunning known successful attacks and edge cases on every release so safety regressions block promotion. Since CI testing is preventive validation rather than runtime enforcement, the option does not provide the required assurance by itself.
Answer C is incorrect because Amazon Bedrock Guardrails prompt attack filter has value for detecting prompt injection and jailbreak attempts intended to override developer instructions, bypass safeguards, or expose protected instructions. In contrast, this scenario depends on automated adversarial security testing before release.
Answer D is incorrect because Production canary attack probes and monitoring is relevant to sending controlled adversarial probes and watching safety metrics after deployment to catch environment-specific regressions early. Because canary probes detect failures but should not expose real users to unsafe output, the option is technically useful but positioned for a different purpose than the control being requested.
