MosaicLeaks: Data Leakage Through Research AI Agents and the PA-DR Method to Prevent It
Deep research agents that combine private local documents with external tools like web search can leak sensitive information through their queries. A new benchmark, MosaicLeaks, measures this risk, and a training method called PA-DR reduces leakage from 34% to 9.9% while maintaining task performance.
MosaicLeaks is a benchmark created by researchers to study privacy leakage from deep-research agents that interleave private local documents and public web retrieval. It contains 1,001 multi-hop question chains where agents must answer questions that bridge private and public information. Testing showed that agents frequently leak private data through their web queries, and training only for task performance made leakage worse (from 34% to 51.7%). The proposed Privacy-Aware Deep Research (PA-DR) method combines a situational task reward with a learned privacy reward, reducing answer/full-information leakage to 9.9% while keeping strict chain success at 58.7% (near the 59.3% achieved by task-only training). PA-DR also improves sample efficiency, requiring 5-6x fewer generated samples to reach the same task performance as outcome-only reinforcement learning.
- Сокращения
- PA-DR = Privacy-Aware Deep Research — Приватно-осознанное глубокое исследование
- RL = Reinforcement Learning — Обучение с подкреплением
Source: Hugging Face blog —
original