Agent Mosaic: LLM agents leak secrets by combining individually permitted confidential documents
UniversityofSouthernCalifornia · hf · 2026-10-24
USC's Agent Mosaic paper (arXiv:2610.30418) shows that LLM agents frequently leak sensitive conclusions synthesized from multiple confidential documents, even when each individual document was permitted.
Example: from two accessible records — "Participant 4078 received a kit ending in '-B'" and "kit codes ending in '-B' contain placebo" — an agent derives the top-secret conclusion "Participant 4078 is on placebo."
The key insight: access to both sources does not equal permission to reveal the composite, and common safeguards fail to eliminate this mosaic-style disclosure risk.
More from Safety
- Ex-FTC Commissioner Warns AI Firms Could Face Legal Liability for Unleashing Hacking Agents — GaryMarcus · 2026-10-24
- X now requires face verification for new signups, photo sent to AWS — ThePeterMick · 2026-10-24
- NHTSA proposes removing regulatory barriers for driverless cars as France resists Tesla FSD — mitchdeg · 2026-10-24
- Matthew Green: AI labs sounded the alarm, but it still wasn't enough — matthew_d_green · 2026-10-24
- Quant hallucination forensics: fabricated prompt injections, pirate puzzles and a phantom Windows attack — Consumerbot37427 · 2026-10-24
- Etsy cracks down on AI slop listings, bans bot-generated items with no human input — lilyraynyc · 2026-10-24