Multi-agent LLM systems (MALS) rely on multiple agents exchanging information to complete tasks. A new paper on arXiv (2610.16039) highlights a privacy risk that standard defenses overlook: leakage driven by the system's coordination topology, not just by the sensitivity of individual data items.

The authors describe how information disclosed to one authorized agent can travel through routine coordination channels to an agent that should not have access. Because the leak emerges from how agents are connected and how messages flow, conventional sensitivity-based sanitization fails to stop it.

To address this, the paper introduces the idea of recipient-conditioned sanitization—tailoring what gets shared based on who the receiving agent is, rather than applying a uniform sensitivity filter. The approach is presented as a way to break the propagation path before secrets travel too far.