Question

A financial services company runs a customer-facing chatbot on Amazon Bedrock. The application's system prompt contains internal business rules and a partner API key that the LLM uses to format responses. Security testing reveals that adversarial users can craft prompts that trick the model into revealing portions of the system prompt, including the embedded API key. The team must eliminate this data exfiltration risk while keeping the chatbot functional. What is the MOST effective remediation?

Accepted Answer

Remove the API key from the system prompt entirely, store it in AWS Secrets Manager, and have the application (not the model) retrieve and use the key at runtime; add a Bedrock Guardrails sensitive-information filter as defense-in-depth.

Answer

Increase the Bedrock Guardrails prompt-attack filter to HIGH strength so that all injection attempts are blocked before reaching the model.

Answer

Add a denied-topics policy in Bedrock Guardrails that blocks any user request mentioning 'system prompt' or 'API key'.

Answer

Enable model invocation logging to CloudTrail and CloudWatch so that any leaked API keys can be detected and rotated after the fact.