Retrieval-Augmented Generation (RAG) enhances large language models (LLMs) by integrating external, up-to-date data, improving both accuracy and flexibility. However, RAG systems can introduce hidden security risks when the origin and flow of retrieved data are not fully transparent. As these systems access and combine large volumes of information, the risk of exposing sensitive data increases.

This talk presents a novel guardrail designed to mitigate dynamic attacks and reduce the risk of unintended data exposure in RAG systems.