MOSAICLEAKS: CAN YOUR RESEARCH AGENT KEEP A SECRET?

Summary from the TLDR AI newsletter

MosaicLeaks highlights the privacy risks of deep research agents that combine private documents with web retrieval, often leaking sensitive information. The proposed solution, PA-DR, enhances both performance and privacy by reducing answer/full-information leakage from 34% to 9.9% while maintaining task success. PA-DR uses rewards for safe query construction rather than relying on user prompts, making agents perform better without compromising data security.