How to Block Violent Bedrock Responses for Children

AIP-C01 By islever Editorial Last updated Oct 19, 2026

Implement input and output safety controls. Implement prompt engineering strategies and governance for FM interactions.

Short answer

B Apply Amazon Bedrock Guardrails content filters to model responses at high strength with a block action and lower the flow's inference temperature.

Question

A company is designing a GenAI solution to help children between the ages of 8 and 14 understand world history. The system must support concurrent use for up to 30 children in a classroom and provide responses in less than 1 second. The company is using Amazon Bedrock and Amazon Nova Pro with a knowledge base that contains history textbooks that are written for students between the ages of 6 and 21. The company uses an Amazon Bedrock flow to ensure that the solution uses the knowledge base to generate every response. The company sets the prompt response temperature to 0.7.
Early testing of the solution results in some responses that are historically inaccurate, violent, and not appropriate for children. A GenAI developer must prevent the solution from responding to prompts inaccurately and inappropriately.
Which solution will meet this requirement?

Increase the inference temperature in the flow. Add a node to the flow after the initial response that calls the Amazon Nova model to remove any violence from the response. Return the result of the model call.
Configure Amazon Bedrock guardrails for the model responses by using content filters. Set the content filter strength high and set the guardrail action to block. Add the guardrail to the flow. Reduce the inference temperature in the flow. Correct Answer
Use Amazon Bedrock Data Automation with a custom output that removes all violent content from the knowledge base. Increase the inference temperature in the flow.
Configure Amazon Bedrock guardrails for the prompts by using content filters. Set the content filter strength to high and set the guardrail action to block. Add the guardrail to the flow. Reduce the inference temperature in the flow.

Key Takeaways

The unsafe content is being produced by the model in its output, so output-side guardrails plus a lower temperature are required — the trap is filtering only the prompts or bolting on a second corrective model call.

The scenario tests how to stop Amazon Bedrock from producing violent, child-inappropriate and historically inaccurate answers in an Amazon Nova Pro knowledge-base flow. Applying Amazon Bedrock Guardrails content filters to the model responses at high strength with a block action, and lowering the flow's inference temperature, is the configuration that meets the requirement.

Common mistake: Option D is the most common wrong pick: learners assume filtering the incoming prompts blocks the problem, but the violent and inaccurate material is generated by the model itself, so prompt-only guardrails never see it. Option A is also tempting because it adds a cleanup model call, but raising the temperature worsens the problem and the extra Nova call breaks the sub-second latency target.

Expert Analysis

Why the Answer Is Correct

Option B applies Amazon Bedrock Guardrails content filters directly to the model responses, which is where the violent and inappropriate content is actually being emitted. Setting filter strength to High with the guardrail action set to Block guarantees those responses never reach the children, and attaching the guardrail to the Bedrock flow enforces it on every path through the solution. Reducing the inference temperature from 0.7 lowers token randomness, which reduces hallucinated historical inaccuracies and over-dramatized output. Together, output filtering and a lower temperature address both the safety and the accuracy defects with no added latency.

Why the Other Options Are Wrong

Option D targets only the prompt (input) side, so model-generated violent or fabricated history passes through untouched because the user prompts themselves may be perfectly benign. Option A raises temperature, which increases randomness and makes both inaccuracy and unsafe elaboration worse, while the second Nova call adds round-trip latency that jeopardizes the stated sub-second response requirement. Option C misuses Amazon Bedrock Data Automation, which is built for document and media content extraction rather than filtering a knowledge base at response time, and it again increases the temperature. Neither A nor C enforces an actual policy-based block on unsafe model output.

Community Comment Notes

The single commenter, TNT87, selected option B, quoting the reasoning as "Configure Amazon Bedrock guardrails for the model responses by using content filters" with a high strength and block action. That reasoning matches the requirement because the offending content appears in the generated responses, not in the students' questions. The vote data is unavailable here, so the verdict rests on the technical reading of where the unsafe content originates.

Official Reference

https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails.html

This link opens the official vendor documentation in a new tab.

Exam Strategy

When unsafe or inaccurate content appears in generated outputs, choose guardrails applied to model responses rather than prompts, and pair them with a lower temperature. Reject any option that raises temperature or adds an extra model call when a sub-second latency budget is stated.

Frequently Asked Questions

Why filter model responses instead of prompts when the content is violent?

The violent and inaccurate text is generated by Nova Pro during inference, not typed by the children. Only output-side content filters can inspect and block that generated content.

Does lowering temperature alone prevent inappropriate answers?

No. A lower temperature reduces randomness and historical hallucination but never guarantees blocking. You still need guardrails set to High strength with the block action.

More AIP-C01 FAQ →

Discussion 1 contributions

Compiled from public community threads and reader submissions, lightly edited for clarity. Handles and vote tallies are shown as recorded, and comments reflect the views of individual contributors.

From public community threads 1

TNT87 👍 1 Answered B

Configure Amazon Bedrock guardrails for the model responses by using content filters. Set the content filter strength high and set the guardrail action to block. Add the guardrail to the flow. Reduce the inference temperature in the flow.

Reader comments & corrections

No reader corrections yet — spotted an error or have a note? Share it below.

Practice All AIP-C01 Questions

Access 110 questions with complete answers and detailed explanations.

View Full AIP-C01 Practice Test →

← Back to AIP-C01 Study Guide