MosaicLeaks Exposes Privacy Leakage in Research Agents — and a Training-Based Fix

Summary

ServiceNow Research published MosaicLeaks, a benchmark and training method exposing a privacy failure mode in research agents that perform multi-hop reasoning across private enterprise documents and the public web. The core risk is the "mosaic effect": individually innocuous outbound web queries, observed together by an adversary monitoring the agent's traffic, can reassemble private enterprise facts the adversary never directly saw. The work defines three leakage levels — intent leakage (inferr