Federated guard reduces multi-agent attack success in tests

Researchers introduced FGLGuard, a federated graph-learning safety guard for LLM multi-agent systems, in an arXiv paper. The paper says operators train on private judge-labeled episode graphs and share model updates rather than traces. Across three benchmarks, it reports outperforming an in-domain centralized ceiling; in a live AgentDojo test, it cut ground-truth attack success by 43% with near-unguarded utility. The results are preprint findings, not an independently verified deployment.