XG-Guard Enhances Multi-Agent Security

机器之心 · wechat · 2026-07-10

This article introduces XG-Guard, an unsupervised security framework designed to defend against attacks on LLM multi-agent systems. Addressing the coarse granularity and lack of interpretability in existing methods, it proposes dual-level representation encoding, topic-based anomaly detection, and the fusion of coarse and fine-grained scores to identify and explain anomalous Agents.

Original post →

More from Safety

Safety channel →