Agent Mosaic: Recognizing Confidentiality Leaks in Agentic LLM Systems

One-line takeaway

The paper quantifies the agent "mosaic effect" of assembling sensitive conclusions from multiple documents: even under benign queries the best model leaks forbidden composites in 7.08% of simulated answers, and with classification headers and aggregation rules 32.37% of answers remain unsafe.

论文量化了智能体的「马赛克泄密」：从多份可访问文档拼出敏感结论。即使是良性提问，表现最好的模型在模拟中仍有 7.08% 的回答泄露禁止的组合信息；加分级标头和聚合规则后，仍有 32.37% 的模拟回答不安全。