RAG Research Advances Across Domains as ‘TabooRAG’ Exposes Transferable Blocking Risk

Overview

• Researchers introduce TabooRAG, a strict black-box blocking attack that injects retrievable documents to trigger refusals, reaching up to 96% success on GPT-5.2 across seven LLMs and three datasets.
• TabooRAG attributes cross-model transferability to alignment homogeneity, finding that overlapping refusal criteria create a shared attack surface for RAG systems.
• MA-RAG, a multi-round agentic framework for medical question answering, iteratively reconciles conflicting responses with targeted retrieval and averages a +6.8 point accuracy gain over the backbone model on seven benchmarks.
• A proposed multi-agent RAG system for state transportation departments adds specialized retrieval, generation, evaluation, and query-refinement agents and indexes technical figures by converting them to searchable text with a vision-language model.
• A legal benchmark on LaborBench reports STARA at 83% accuracy versus 58% for Westlaw AI and 64% for Lexis+ AI, with error analysis indicating STARA may effectively reach 92% due to omissions in the DOL attorneys’ ground truth and highlighting persistent retrieval and reasoning failures.