AI & LLM Security
RAG & Vector Database Attacks
Poison the knowledge base and exploit retrieval-augmented pipelines.
Retrieval-Augmented Generation (RAG) gives an LLM external knowledge: user query → embed → search a vector database → stuff results into the prompt → generate. Every stage is attackable.
- Indirect prompt injection via documents — plant instructions in a file that gets ingested; when retrieved, the model obeys them.
- Knowledge-base poisoning — inject malicious/false content so the model returns attacker-controlled answers.
- Cross-tenant leakage — weak isolation lets one user's query retrieve another tenant's documents.
- Embedding inversion / data exposure — sensitive data recoverable from stored vectors or over-broad retrieval.