MosaicLeaks: Can your research agent keep a secret?

MosaicLeaks reveals deep research agents risk privacy by leaking sensitive information through external queries, even when combining private documents with public tools. A new task with multi-hop questions shows agents frequently leak private data, and training solely for performance worsens this. The proposed Privacy-Aware Deep Research (PA-DR) method improves task success to 58.7% while significantly reducing answer/full-information leakage from 34.0% to 9.9% by training agents to avoid revealing private facts through query logs.