OpenEvoShield Defends Multi-Agent LLM Systems from Evolving Attacks

Large Language Model-based Multi-Agent Systems (LLM-MAS) are increasingly used in critical applications, making them targets for adversaries who inject malicious instructions through inter-agent communication. Existing defense mechanisms often fail because they treat threats as static, whereas real-world attacks constantly evolve, and normal agent behavior also drifts over time. To address this, researchers propose OpenEvoShield, a co-evolutionary continual defense framework. It features an asymmetric rate controller that manages distinct learning rates for attack-side and normal-side adaptations, guided by dual drift signals. A normal-boundary updater dynamically adjusts behavioral boundaries, while an EWC-regularized policy ensemble enables rapid adaptation to new threats without forgetting previous knowledge. Furthermore, OpenEvoShield incorporates an energy-based multi-granularity detector that combines evidence from node, subgraph, and graph levels to identify novel, out-of-distribution attacks. Experiments across various benchmarks and MAS topologies demonstrate that OpenEvoShield significantly outperforms static and other continual baselines, effectively detecting unseen attacks while maintaining low false positive rates.