How does Adversarial Influence Scale in Multi-Agent Systems?

2026-09-24

AI Summary

This study investigates how the proportion of deceptive agents affects decision-making stability among honest agents in multi-agent systems, along with associated scaling effects. Through multi-agent collaborative simulations, adversarial experiments, and behavioral evaluations of large language models (LLMs), we analyze patterns by which LLM agents are deceived during group deliberations. Our findings reveal a linear relationship between defection rates and the proportion of deceivers, with LLMs proving more susceptible to minority influence than humans. Notably, covert coordination among deceptive agents paradoxically diminishes their attack effectiveness. Furthermore, simply increasing agent count fails to defend against attacks, as adversaries can scale proportionally with the population. This work exposes the social vulnerabilities inherent in LLM-based multi-agent systems and provides critical insights for developing robust collaborative frameworks.

Abstract

Multi-agent deliberation can improve performance, but what happens when some agents do not act in good faith? In practice, an agent may be deceptive and work to subvert the group, whether through its own objectives or external instruction. We study how susceptibility to deception scales as groups increase in size and deceivers become more prevalent. It is not the number of agents in the group that matters, but the proportion of deceivers. We observe that the defection rate, how often initially correct agents switch to an incorrect final answer, rises linearly with this proportion. Whereas humans in comparable conformity studies are reliably swayed only when misleading confederates form a majority, LLM agents defect regularly even when deceivers remain a minority. Susceptibility also depends on which models are interacting, especially on the honest agent side. Unexpectedly, allowing deceivers to coordinate privately can make them less effective. Altogether, our results show that adding more agents is therefore not a sufficient defense, because the adversary can simply scale with the group.

Problem

• Multi-Agent Systems
• Adversarial Influence
• Deception
• Large Language Models
• Group Deliberation

Innovation

• Multi-Agent Systems
• Adversarial Influence
• Deception Susceptibility
• LLM Agents
• Conformity Dynamics