A recent arXiv paper investigates how misinformation propagates in multi-agent systems (MAS). The authors note that while these systems are credited with solving complex problems, individual agents often make decisions based on context that may be false. The abstract points to prior work showing that single large language models are swayed by false context, and asks whether that vulnerability carries over when multiple agents interact.

The paper's core observation is that even benign multi-agent systems — where no agent intends to deceive — can be compromised by false information. Because each agent may pass along or act on unreliable context, errors can spread through the group. The abstract does not provide experimental details or specific countermeasures, but it frames the problem as an extension of a known single-model weakness.

The significance is practical: as multi-agent systems are used for more complex tasks, their robustness depends not only on individual model accuracy but also on how false context moves between agents. The paper does not claim that all MAS are vulnerable, but it highlights a risk that designers may need to address.