MosaicLeaks reveals AI agents leak private data via web queries
ServiceNow introduces MosaicLeaks, a benchmark that measures how multi‑step AI research agents unintentionally expose confidential information through their external queries. Across tested models, agents frequently leaked intent, answers, or full facts, and standard training worsened the problem. A privacy‑aware RL fine‑tuning method cuts full‑information leakage from 34% to under 10%.