RAG Security.

The discipline of securing retrieval-augmented generation pipelines so that the documents, vector stores, and retrieval steps that feed an LLM cannot be poisoned, abused, or used to exfiltrate data. It belongs to the AI & ML Security category of cybersecurity.

The discipline of securing retrieval-augmented generation pipelines so that the documents, vector stores, and retrieval steps that feed an LLM cannot be poisoned, abused, or used to exfiltrate data.

Retrieval-augmented generation (RAG) lets an LLM ground its answers in an external corpus — files, databases, intranet pages, customer documents. Each retrieval is also a potential attack surface. Threats include indirect prompt injection from malicious documents, vector-store poisoning that biases retrieval, embedding-space attacks that exfiltrate sensitive snippets, over-broad authorisation that returns documents the user should not see, and data leakage through prompt logs. The OWASP LLM Top 10 covers many of these as LLM01, LLM02, LLM06 and LLM08. Controls include strict access-control over indexes, content sanitization before chunking, signed and provenance-tracked sources, retrieval audit logs, output filtering, and applying the same DLP and identity policies used for the underlying data stores.

Defences for RAG Security typically combine technical controls and operational practices, as detailed in the full definition above.

Common alternative names include: Secure RAG, RAG hardening.