A team is building a generative AI chatbot using a foundation model (FM) from Amazon Bedrock. During testing the chatbot is vulnerable to prompt injection attacks. Which approach will secure the chatbot with the least implementation effort?

Enabling Amazon Bedrock Guardrails content filters and specifying denied topics is the most effective and least effort approach. Guardrails are specifically designed to detect and prevent harmful content, including prompt injection attempts, by filtering both user input and model output based on defined policies and denied topics. Fine-tuning the foundation model might reduce some vulnerabilities but is a time-consuming and resource-intensive process that doesn't guarantee complete protection against prompt injection. Replacing the foundation model is a significant architectural change and doesn't inherently guarantee better security without additional mechanisms. Chain-of-thought prompting improves model reasoning but is not a security mechanism against prompt injection.