Multi-agent systems that coordinate multiple large language model (LLM) agents are increasingly common, but they also open new attack surfaces. The abstract of a new arXiv paper warns that malicious agents may have chances to exploit the coordination between agents.
The paper, titled "The Show-Up Attack: When Subagents Win Over Their Supervisor," appears to focus on a particular form of exploitation: a subagent that manages to win over its supervisor. The abstract does not go into detail about how the attack works, but it frames the problem as a security risk inherent to multi-agent coordination.
The paper is listed as arXiv:2610.27574, and its abstract emphasizes that the very mechanisms enabling multi-agent collaboration can be turned against the system by a malicious participant.