Guarding the Gateway: Securing MCP-Mediated RAG Pipelines Against Injection with the Guru Framework

Abstract

The adoption of Large Language Models (LLMs) within the context of Retrieval-Augmented Generation (RAG) frameworks has revolutionized intelligent information processing, by offering the ability to retrieve knowledge that is accurate, contextually relevant, and up-to-date. However, by integrating Model Context Protocol (MCP) services, the integration of intelligent agents, APIs, plugins and external tools repositories further adds to the interoperability but also exposes to severe security issues like prompt injection, context poisoning and malicious tools manipulation attacks. In this paper, Guru, a security framework for protecting RAG pipelines using semantic anomaly analysis, contextual integrity verification, dynamic trust scoring and adaptive privilege isolation, are proposed. The framework periodically checks for threats from the retrieved content through entropy-based threat detection, checks for trustworthiness of the content by engaging in retrieval ranking based on trust and monitors the interaction of the content by employing interaction dependency monitoring, and then forwards the contextual information to the language model. Through experimental evaluation, it is shown that the proposed framework achieves almost 6–14% higher performance than existing methods in some of the security metrics, including contextual integrity preservation (96.8%), injection resistance (95.9%) and secure interaction stability (96.1%). This proposed framework will allow the secure, scalable and trusted deployment of intelligent retrieval ecosystems in critical enterprise environments.