Surveying the RAG Attack Surface and Defenses: Protecting Sensitive Company Data

As the performance of Large Language Models (LLMs) improves, they are quickly becoming a part of daily life for many people. Even industry companies are adopting them as tools to increase employee productivity. However, due to the huge computational requirements to train an LLM, companies need methods to customize their LLMs without training their own or retraining a pre-trained model. Therefore, retrieval-based methods, such as Retrieval-Augmented Generation (RAG) are becoming popular to reduce hallucinations and incorporate application-specific information into LLM generations without fine-tuning or training. However, the integration of a company’s sensitive data into LLM generations using RAG changes the LLM’s attack surface. Current literature lacks a complete view of the LLM’s changing attack surface due to RAG. While many authors have noted specific vulnerabilities, a survey of the attack surface is not present. This work is a survey on the vulnerabilities introduced to LLMs by using RAG and how these vulnerabilities present security risks to industry companies utilizing RAG to incorporate their sensitive data into LLM generations. This work then provides a survey of the state-of-the-art defense methods against these vulnerabilities.