SafeRAG introduces a comprehensive benchmark for evaluating security vulnerabilities in Retrieval-Augmented Generation systems, revealing critical attack vectors through external knowledge manipulation.

• Classifies and evaluates multiple attack vectors targeting the RAG pipeline
• Demonstrates how attackers can compromise LLMs by manipulating knowledge sources
• Provides a structured framework to assess and improve RAG security posture
• Highlights the security tradeoffs between knowledge enhancement and vulnerability exposure

This research is crucial for organizations deploying RAG systems in production environments, as it exposes security risks that could lead to misinformation, data leakage, or system manipulation when incorporating external knowledge sources into LLMs.

SafeRAG: Benchmarking Security in Retrieval-Augmented Generation of Large Language Model