A Security Meta-Model for Retrieval-Augmented Generation Systems
A RAG security meta-model catalogs attacks, weaknesses, risks, and CIA impact drawn from 43 papers.
The paper introduces a security meta-model for retrieval-augmented generation that relates pipeline surfaces, attacks, weaknesses, risks, and confidentiality, integrity, and availability impact. It was built from an iterative review of 43 publications from 2023 to 2026 and instantiated as a catalog of reported threats and remediations. Filtering that catalog by a deployment configuration produces an applicable risk profile, and a web visualizer exposes causal chains and stakeholder views. The authors report more attack-focused than defense-focused literature, threat concentration at ingestion, and gaps in output-integrity coverage, assessed against the OWASP LLM Top 10 for textual, graph, and multimodal RAG.
- Meta-model connects RAG surfaces, attacks, weaknesses, risks, and CIA impact.
- Catalog is drawn from 43 publications published from 2023 through 2026.
- Filtering by deployment yields a configuration-specific risk profile and graph view.
- Authors report more attack research than defenses and a concentration of ingestion threats.
- Coverage is compared with the OWASP LLM Top 10 across RAG configurations.
From the original article
1 min read at arxiv.orgRetrieval-Augmented Generation (RAG) systems extend large language models (LLMs) with external knowledge through a multi-stage pipeline. While this architecture can improve the factual grounding of generated answers, it introduces structural attack surfaces that extend beyond those of standalone LLMs.
By Steve Nouyep · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.