# MosaicLeaks: Can your research agent keep a secret?

DevFeed: [MosaicLeaks: Can your research agent keep a secret?](<https://devfeed.tech/articles/mosaicleaks-can-your-research-agent-keep-a-secret-7054.md>)

Original publisher: [Read original article](<https://huggingface.co/blog/ServiceNow/mosaicleaks>)

Author: Alexander Gurung; Rafael Pardinas

Published: 2026-06-18T18:13:13Z

Content type: article

Language: en

Sources: [Hugging Face - Blog](<https://devfeed.tech/sources/hugging-face-blog.md>)

Topics: [Security](<https://devfeed.tech/topics/security.md>), [Machine Learning, Security Attacks](<https://devfeed.tech/topics/machine-learning-security-attacks.md>)

Tags: [agent](<https://devfeed.tech/tags/agent.md>), [agents](<https://devfeed.tech/tags/agents.md>), [cloud](<https://devfeed.tech/tags/cloud.md>), [external](<https://devfeed.tech/tags/external.md>), [healthcare](<https://devfeed.tech/tags/healthcare.md>), [infrastructure](<https://devfeed.tech/tags/infrastructure.md>), [local](<https://devfeed.tech/tags/local.md>), [migration](<https://devfeed.tech/tags/migration.md>), [models](<https://devfeed.tech/tags/models.md>), [privacy](<https://devfeed.tech/tags/privacy.md>), [research](<https://devfeed.tech/tags/research.md>), [retrieval](<https://devfeed.tech/tags/retrieval.md>), [rl](<https://devfeed.tech/tags/rl.md>), [security](<https://devfeed.tech/tags/security.md>), [tools](<https://devfeed.tech/tags/tools.md>), [training](<https://devfeed.tech/tags/training.md>)

## AI overview

MosaicLeaks examine how deep-research agents can expose private enterprise information through their external web queries. The proposed Privacy-Aware Deep Research training method improves strict multi-hop task success while substantially reducing answer and full-information leakage.

## Source excerpt

Deep research agents increasingly combine private local documents with external tools like web retrieval, creating a privacy risk: an agent's external queries may leak sensitive information. MosaicLeaks proposes a new deep-research task with multi-hop questions that interleave public and private information. Across the models we tested, agents frequently leaked private information, and training only for task performance made it worse.