MosaicLeaks: Can your research agent keep a secret?

Ignoring the mosaic effect allows external observers to reconstruct confidential business strategies by aggregating fragmented search logs. When agents interleave private documents with public web tools, they often inadvertently broadcast internal milestones or security disclosures. Organizations that successfully implement privacy-aware training can deploy agents for complex competitive analysis without compromising their proprietary data. However, these models still face limitations in balancing high-performance reasoning with the strict obfuscation of sensitive query parameters.

Alexander Gurung and Rafael Pardinas of ServiceNow developed the MosaicLeaks framework to quantify how research agents leak information through outbound traffic. The team introduced Privacy-Aware Deep Research (PA-DR) training to mitigate these risks, demonstrating that performance-only training actually exacerbates leakage. While PA-DR improves reasoning accuracy, the approach remains constrained by the inherent difficulty of sanitizing queries that rely on private context for downstream tasks.

This week

• Audit current research agent workflows to identify endpoints where internal documents are accessed alongside external search tools.
• Review outbound traffic logs for recurring patterns that could reveal specific project goals or proprietary metrics.

This quarter

• Implement privacy-aware training constraints to ensure agent queries do not carry forward sensitive bridge entities.
• Establish a red-teaming protocol to test whether agents leak internal milestones during routine market research tasks.

This year

• Transition to agent architectures that utilize local-only processing for sensitive data analysis.
• Standardize privacy-leakage benchmarks for all internal AI tools to ensure compliance with data protection requirements.