Secure RAG applications using prompt engineering on Amazon Bedrock

This article discusses how to secure Retrieval Augmented Generation (RAG) applications that use large language models on Amazon Bedrock. It covers different prompt-level security threats like prompt injection, prompt leaking, and jailbreaking, and proposes guardrails and best practices to mitigate these threats.

Specifically, the article covers:

• Introduction to LLMs and RAG
• Different security layers of generative AI solutions
• Common prompt-level threats
• Proposed prompt template guardrails (e.g. salted tags, threat detection instructions)
• Comparison of guardrail template vs. basic template on security benchmark
• Key takeaways and learnings
• Conclusion