Judging Before Propagation: Task-Graph Contextual Integrity for Multi-Agent Privacy.

arXiv:2610.39532v1 Announce Type: cross Abstract: Multi-agent LLM systems increasingly rely on persistent memory and inter-agent communication. These capabilities create privacy risks that extend beyond individual interactions: a release appropriate in isolati