Seven security challenges in cross-domain multi-agent LLM systems
摘要
Large language models are evolving into autonomous agents that collaborate across organizations for tasks like disaster response and supply-chain optimization. However, such cooperation breaks unified trust assumptions: a benign agent may leak secrets or violate policy when interacting with untrusted peers. This paper maps the security agenda for cross-domain multi-agent LLM systems, introducing seven categories of novel challenges alongside plausible attacks, evaluation metrics, and research directions.