VectorSmuggle: Steganographic Exfiltration Attack on RAG Vector Databases

arXiv·high signal

New paper demonstrates a class of steganographic exfiltration attacks against RAG systems: an attacker with write access to the embedding ingestion pipeline can encode arbitrary data into high-dimensional vectors that evade distributional checks. Major vector-store products lack native embedding integrity controls, ingestion-time anomaly detection, or cryptographic provenance attestation. The authors propose a cryptographic provenance defense as countermeasure. Directly relevant to anyone deploying production RAG.