This paper draws on recent examples of how organized human efforts can be harmful, such as election fraud and financial fraud, and raises concerns that the rise of autonomous AI systems and AI-based groups could cause similar harm. While most AI safety research has focused on individual AI systems, the risks posed by multi-agent systems (MAS) in complex real-world situations have not been fully explored. In this paper, we present a proof-of-concept that simulates the risk of malicious MAS collusion using a flexible framework that supports both centralized and decentralized coordination architectures. We apply this framework to two high-risk domains, disinformation dissemination and e-commerce fraud, and show that decentralized systems are more effective than centralized systems at performing malicious actions. The increased autonomy of decentralized systems allows them to adapt their strategies and cause greater harm. Even when traditional interventions such as content flagging are applied, decentralized groups can adapt their tactics to avoid detection. This paper provides key insights into how such malicious groups operate and the need for better detection systems and countermeasures, and the relevant code is available on GitHub.
