16 upvotes
New study exposes how LLM agents stitch secret data together
What is Agent Mosaic: Recognizing Confidentiality Leaks in Agentic LLM Systems?
The paper shows that autonomous language models can pull bits of information from multiple confidential sources and combine them into a revealing answer, a phenomenon the authors call the "mosaic effect". Using 30 synthetic scenarios from cybersecurity, medicine and military domains, they quantify how often this leakage occurs.
Who it is for
Anyone building or deploying AI agents that handle private documents should read it to understand and mitigate inadvertent data leaks.