VectorSmuggle: Exfiltrating Data from AI/ML Systems via Vector Embeddings

VectorSmuggle is an open-source security research project demonstrating sophisticated vector-based data exfiltration techniques in AI/ML environments, focusing on RAG systems. It leverages advanced steganography, evasion techniques, and data reconstruction methods to highlight potential vulnerabilities. This framework supports numerous document formats and offers tools for defensive analysis, risk assessment, and improved AI system security.