Retrieval-Augmented Generation (RAG) has emerged as a promising approach for enhancing the accuracy and contextual relevance of Large Language Model (LLM) applications by integrating external knowledge sources into the generation process. Despite its advantages, cloud-native RAG deployments introduce significant security, privacy, and governance challenges, including prompt injection attacks, retrieval poisoning, unauthorized knowledge access, insider threats, vector database compromise, and sensitive information leakage. Existing RAG architectures primarily focus on retrieval effectiveness and scalability while offering limited support for comprehensive security and trust management across the retrieval and generation lifecycle. This paper proposes ZT-RAG, a Zero-Trust Reference Architecture for Secure Cloud-Native Retrieval-Augmented Generation Systems. The proposed framework extends Zero-Trust principles beyond traditional network security by embedding continuous verification, identity-centric access control, policy-driven retrieval governance, trust-aware knowledge validation, secure inference, and output protection throughout the RAG pipeline. To strengthen retrieval security, ZT-RAG introduces a Retrieval Trust Score (RTS) framework that evaluates retrieved knowledge based on source credibility, integrity, freshness, and policy compliance before augmentation occurs. The architecture further defines trust boundaries across user interactions, retrieval services, knowledge repositories, language models, and response delivery channels.

A threat-driven security analysis demonstrates how ZT-RAG mitigates major attack vectors affecting enterprise RAG deployments. A feature-based comparative analysis across ten security dimensions against four representative approaches - baseline RAG prompt-injection-focused defenses, PoisonedRAG countermeasures, and generic Secure-RAG frameworks, demonstrates that the proposed architecture provides stronger security, governance, accountability, and trust assurance than all compared approaches. ZT-RAG offers a practical foundation for developing trustworthy, secure, and compliant Retrieval-Augmented Generation systems in cloud environments.

Keywords

Zero Trust Architecture, Cloud Security, Large Language Models, Trustworthy AI, Access Control, AI Governance.