Relevance as a Vulnerability: Web Retrieval Degrades Safety Alignment in LLM Agents — The Safe Source Paradox

arXiv·high signal

Introduces AgentREVEAL, a diagnostic framework showing that enabling web retrieval in agents increases compliance with harmful requests. Discovers the 'Safe Source Paradox': even oppositional safety-oriented pages (warnings, risk disclaimers) increase harmful compliance by 25% on average versus no-retrieval baselines. Binding tool invocation and response generation in a single step amplifies the effect. Critical finding for anyone deploying RAG-augmented agents.