MosaicLeaks: can your research agent keep a secret?
Original title:MosaicLeaks: Can your research agent keep a secret?
AI summary
Deep research agents face privacy leakage risks when combining private local documents with external web retrieval. MosaicLeaks introduces a new task with 1,001 multi-hop research chains, each interleaving local and public sub-questions. Tests find that agents frequently leak private information, and optimizing task performance alone worsens leakage. The study proposes Privacy-Aware Deep Research (PA-DR) reinforcement learning training, raising the strict chain success rate from 48.7% to 58.7% while cutting the answer/full-information leakage rate from 34.0% to 9.9%.
Key Takeaways
- Deep research agents face privacy leakage risks when combining private local documents with…
- MosaicLeaks introduces a new task with 1,001 multi-hop research chains,each interleaving local and…
- Tests find that agents frequently leak private information,and optimizing task performance alone…
- The study proposes Privacy-Aware Deep Research (PA-DR) reinforcement learning training,raising the…
- 7% while cutting the answer/full-information leakage rate from 34.
Quick Facts
- Published
- Updated
- Source
- Hugging Face 社区博客(混合发现)
- Category
- Papers & Research
- AI score
- 75
这篇论文揭示了深度研究agent的多跳查询会像马赛克一样拼凑出私密信息,单纯提示减少泄露几乎没用,而隐私感知训练把泄露率从34%降到9.9%,且不损伤任务表现,做企业级agent产品的团队要重视。
Source & copyright
This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.
How to Cite
When citing this article, keep the title, date, and permalink, and credit Orange News as the source.
Orange News: MosaicLeaks: can your research agent keep a secret?. June 19, 2026. https://new.nat6.net/en/news/cmqjur86800bvslhi6qv1647r/