LLM-Based Multi-Agent Systems Vulnerable to Cooperative Attacks

A new research paper on arXiv (2605.28104) identifies a critical vulnerability in Large Language Model-based Multi-Agent Systems (MAS): cooperative attacks by malicious agents. Previous defense strategies assumed attackers act independently, but the paper argues that coordinated malicious agents can share information and dynamically adjust strategies through multi-round interactions, enabling more effective attacks. To counter this, the authors propose STAR (Sentence-Level Trustworthiness Analysis and Rectification), a defense mechanism that evaluates and corrects suspicious statements at the sentence level. The work highlights a growing need for robust security in collaborative AI systems.