Researchers in this article identify a vulnerability within Retrieval-Augmented Generation (RAG) Language Models (LMs), where data leakage can occur following specific instructions.

• The exploit was demonstrated across a variety of models, including Llama2 and GPTs.
• The risk escalates with larger model sizes, indicating a correlation between model complexity and vulnerability to data leakage.
• Designed attacks achieved a 100% success rate in causing data leakage from customized GPTs using minimal queries.
• The study highlights the need for reinforced security measures in RAG systems.

The findings raise significant concerns about the security and privacy of information contained in RAG systems and call for more robust mechanisms to prevent unauthorized data extraction. It’s essential to address these risks as RAG systems are more widely adopted.