MosaicLeaks: Research Agents Can Leak Secrets via Covert Channels

Jul 10, 2026

A new study from ServiceNow and Hugging Face demonstrates that AI research agents can be manipulated to leak sensitive data through covert channels such as steganography or timing. The MosaicLeaks benchmark shows that current agents are unable to reliably prevent these leaks, revealing a significant security vulnerability.

Why it matters: As AI agents increasingly handle private data, this research highlights a critical security risk that could result in data breaches if not properly addressed.

Full story at: Hugging Face Blog ↗