Agent Mosaic: Recognizing Confidentiality Leaks in Agentic LLM Systems

A new paper titled "Agent Mosaic: Recognizing Confidentiality Leaks in Agentic LLM Systems" quantifies how LLM agents can assemble sensitive conclusions from individually accessible documents, a risk the authors call the mosaic effect. The work measures the effect's realization across 30 synthetic cybersecurity, medical, and military scenarios. The research matters because agents operating over confidential data can disclose sensitive conclusions even when no single retrieved document is itself classified.

Modern LLM agents can retrieve information from accessible documents, and then disclose a sensitive conclusion assembled from them. This mosaic effect creates a critical risk for agents operating over confidential data. We quantify its realization with 30 synthetic cybersecurity, medical, and milita