MosaicLeaks reveals privacy vulnerabilities in deep research agents that combine private documents and web searches, leading to potential leakage of sensitive information. The proposed Privacy-Aware Deep Research (PA-DR) method improves task accuracy and decreases information leakage significantly, from 34.0% to 9.9% for full-information leakage.
MosaicLeaks highlights the privacy risks in deep research agents that merge private local documents with external web queries. The process can inadvertently reveal sensitive information, as illustrated by a scenario where an agent's innocuous searches could collectively disclose confidential details about a corporate cloud migration.
The term 'mosaic effect' describes how separate routine queries can reconstruct sensitive corporate information when viewed together. An observer monitoring query logs may piece together private facts that are otherwise secure within internal documents.
MosaicLeaks measures potential information leakage in three escalating levels: Intent leakage, which reveals what the agent is researching; Answer leakage, where the query log provides sufficient information to answer private questions; and Full-information leakage, where the observer can identify and confirm private facts without explicit guidance.
To combat these risks, researchers have developed the PA-DR training method, which focuses on minimizing information leakage while improving the correct response rate of research queries. This method raised the success rate of providing correct chain responses from 48.7% to 58.7% and reduced full-information leakage significantly from 34.0% to 9.9%.
This research is vital as it addresses the pressing challenge of maintaining privacy in AI systems that require access to both public and private information. The findings suggest that without adequate safeguards, deep research agents could pose substantial risks to sensitive enterprise data.
✨ This summary was generated by AI from the outlets' reporting listed below. It is not independently verified and may contain errors — check the original sources. How BrevFeed works →
MosaicLeaks reveals privacy vulnerabilities in deep research agents that combine private documents and web searches, leading to potential leakage of sensitive information. The proposed Privacy-Aware Deep Research (PA-DR) method improves task accuracy and decreases information leakage significantly, from 34.0% to 9.9% for full-information leakage.