Analysis · curated 15 Jul 2026
RAG Security Guide: Prevent Enterprise AI Data Leakage
First reported · updated · 4 reports medium.com
Coverage timeline
Why it matters
RAG pipelines are the architecture most enterprises reach for first, and this guide surfaces recurring weaknesses—especially permission leakage that appears in a majority of deployments—that defenders must address across every stage rather than just hardening the model.
The bdemerson.com guide "RAG Security: Where Retrieval Pipelines Fail" walks through five failure classes in retrieval-augmented generation deployments: poisoned corpora, indirect prompt injection carried in retrieved documents, permission leakage across the index, embedding inversion that reconstructs text from stored vectors, and weak tenant isolation in vector stores. It frames the retrieval pipeline (ingestion, embedding, storage, retrieval, generation) as the real attack surface and offers defensive evaluation habits, referencing research such as the arXiv paper "Pirates of the RAG" on black-box knowledge-base leakage.