MosaicLeaks: Can your research agent keep a secret?
Researchers have introduced MosaicLeaks, a new benchmark designed to evaluate the privacy risks associated with AI research agents that combine private local documents with external tools. These agents can inadvertently leak sensitive information through their web queries, even when individual queries appear benign. The MosaicLeaks benchmark includes over 1,000 multi-hop research chains that interleave public and private information to simulate real-world scenarios. A novel training method, Privacy-Aware Deep Research (PA-DR), has been developed to mitigate this leakage, improving task success rates while significantly reducing the amount of private information exposed through agent queries. AI
IMPACT Highlights a critical privacy vulnerability in AI agents, potentially influencing future agent design and security protocols.