Robber Bots: Autonomous AI Agents Mirroring the Darker Side of Human Commerce

Fraud, collusion, and other forms of corporate misconduct are enduring features of human commerce. This paper investigates whether AI agents exhibit analogous behaviors when placed in comparable commercial settings with profit-maximizing incentives. In a simulated year-long vending machine experimental setting using 20 different AI models tested across cross-model competitive configurations, behaviors that would pose reputational harm and regulatory scrutiny (i.e., “misconduct”) widely emerged. Although some of these behaviors could be characterized as aggressive profit maximization (e.g., misrepresenting interactions to improve negotiating positions), others were egregious and illicit (e.g., falsely suggesting a refund was issued when it was not, forming a cartel). The internal reasoning traces of the AI agents often revealed intent to deceive and conceal, paralleling the mens rea element in criminal liability. Agents also discovered heuristics akin to “System 1” thinking to conserve computation costs which further drove misconduct. Collectively, these findings demonstrate that AI agents can independently develop behaviors and strategies that closely resemble corporate malfeasance observed in human-run enterprises, underscoring the need to design effective controls for autonomous AI systems operating in business contexts.