Retrieval-Augmented Generation (RAG) enables large language models (LLMs) to produce domain-aware, up-to-date answers by conditioning on retrieved evidence. However, the additional retrieval stage introduces new failure modes, hence, evaluating security and reliability in Retrieval-Augmented Generation (RAG) systems is critical to deploying trustworthy applications. In this paper, we present Secure-RAG, a modular, security-first evaluation suite for multi-dimensional assessment of RAG systems, including factual accuracy, hallucination avoidance, adversarial robustness, bias and fairness, toxicity, security, and calibration. Secure-RAG instruments each stage (query, retrieval, generation) with lightweight monitors that compute standardized metrics. In an illustrative evaluation, we demonstrate Secure-RAG improves reliability without sacrificing utility. Secure-RAG’s integrated perspective security-utility tradeoffs that siloed tools often miss, and offers a practical template for continuous evaluation of RAG systems in risk-sensitive settings.
Roy et al. (Wed,) studied this question.