Agent Mosaic Quantifies Confidentiality Leaks in LLM Systems
October 20, 2026
The Agent Mosaic study identifies risks where LLM agents assemble sensitive conclusions from multiple accessible documents. Testing showed that even high-performing models disclosed forbidden composite information in 7.08% of simulated answers.
HOW THIS AFFECTS YOU
●
builderYou must implement classification headers or aggregation rules to prevent information leaks in agentic workflows.
●
policyThis highlights a critical privacy vulnerability in RAG-based agent systems.