Question

A company deploys a customer-facing chatbot built on a foundation model in Amazon Bedrock. During testing, the team finds that some users craft inputs that trick the model into producing offensive language and revealing internal system instructions. Leadership wants a managed capability that filters both harmful user inputs and unsafe model responses without retraining the model. Which approach best meets this requirement?

Accepted Answer

Configure Amazon Bedrock Guardrails to define denied topics, content filters, and prompt-attack detection applied to inputs and outputs

Answer

Increase the temperature parameter so the model produces more varied and less predictable responses

Answer

Fine-tune the foundation model on a labeled dataset of offensive and non-offensive text

Answer

Reduce the max tokens setting so the model cannot generate long harmful responses