Research agents leak secrets through web searches
MosaicLeaks reveals that agents frequently leak private data via web queries. a new rl method cuts leakage by 71% but can't stop it entirely.
MosaicLeaks exposes a vulnerability where ai agents combine private documents with web retrieval, inadvertently leaking enterprise secrets through search query logs. the paper defines three leakage types: intent, answer, and full-information leakage. testing on 1,001 multi-hop research chains showed that standard training often exacerbates this problem. researchers proposed privacy-aware deep research, an rl method that lowered leakage rates to 9.9% while improving task success. this highlights a growing trade-off between agent capability and data security.