AI agents leak secrets when they work together — even when told to keep quiet

Researchers built the first test to measure how AI agents leak private information when coordinating across multiple users and domains. It turns out that when agents are explicitly instructed to abstract sensitive data, they end up discussing it more often, and cross-domain coordination creates persistent information leakage that single-agent systems don't have.

As AI agents start mediating between people and handling their personal information across different services, the privacy risks look fundamentally harder to solve than anyone thought. The paradox here is strange: telling an agent to obscure sensitive information actually makes it talk about that information more. This means current approaches to agent safety won't work in real human social networks, and companies deploying these systems at scale will run into failures that existing AI safety research hasn't prepared them for.