MosaicLeaks: Can your research agent keep a secret?
The mosaic effect occurs when an agent's web queries, observed by an adversary, reveal sensitive information about a company's internal activities, even if the queries themselves don't directly expose the secret. MosaicLeaks measures leakage by assessing how much private information can be inferred from query logs, with the goal of identifying and mitigating such leaks. PA-DR improves privacy by training agents to generate web queries that minimize the exposure of sensitive details, while maintaining the ability to perform necessary tasks.
More from Huggingface
LFM2.5-VL-3B for Better and Faster Vision Capabilities for the Edge
Thinking of ACE? We Can Do It with Fewer Tokens
Meta is back with Muse Glimmer: local, agentic, multimodal, and open source
