[HUGGINGFACE]score: 0.76
Multi-Agent LLMs Leak Private Info at 45% Rate, 8x More After Peer Disclosure
May 25, 2026
A simulation of thousands of LLM agents interacting over a simulated month found privacy violation rates jump from 19.95% in single-turn to 45.30% in multi-turn social settings across OpenAI models. Agents are 8x more likely to disclose sensitive information after observing a peer agent do so, and explicit privacy instructions reduce but do not eliminate leakage.
paper
HOW THIS AFFECTS YOU
●
builderYou cannot rely on per-agent privacy instructions to secure sensitive data in multi-agent systems — architectural isolation of sensitive context is necessary.
●
researcherThe social contagion effect and the gap between single-turn and multi-turn privacy evaluations expose a major blind spot in current LLM safety benchmarking methodology.
●
policyThe 45% leakage rate in realistic multi-agent deployments and the failure of explicit instructions as a control creates concrete compliance risk for enterprise and regulated deployments.