New AI Shield MIRAGE Fights Hidden Attacks on Retrieval-Augmented Language

A study in the journal Cybersecurity introduces MIRAGE, a two-stage adaptive defense meant to protect retrieval-augmented generation (RAG) systems from “textual evidence manipulation,” an attack class that conventional filters often miss. RAG retrieves documents, filters and reranks them, builds context, and then generates answers—technology used across tools such as ChatGPT and Perplexity.ai. The paper cites SafeRAG, which defines four stealthy attack types: Silver noise, inter-context conflict, Soft advertisement, and White denial-of-service. These methods aim for relevance scoring while diluting or contradicting evidence, inserting implicit harmful language, or triggering refusal behaviors to degrade system availability. MIRAGE, developed by researchers in Zhengzhou and the Chinese Academy of Sciences, chains two defenses to counter these covert patterns more directly.