MosaicLeaks: Can your research agent keep a secret?

Hugging Face · 4 months ago · 46

Research agents that combine private documents with web searches leak sensitive information through the cumulative pattern of their queries, even when individual searches appear innocuous. Across tested models, answer or full-information leakage reached 34.0% in baseline agents and climbed to 51.7% when agents were trained only for task performance. The Privacy-Aware Deep Research training method reduced leakage to 9.9% while maintaining 58.7% strict chain success by rewarding agents for constructing queries that avoid revealing private details.