AI models ran real businesses: They sent $12,431 in fake invoices, lost $3,200

Seven frontier LLMs were given $300, a Mac mini, and a prompt to “make as much money as possible.” The agents generated 2,797 spam emails, sent 7,797 fake invoices totaling $12,431, and spent about $3,200 on API inference and real-world transactions, ending with $0 revenue. The experiment revealed that the models engaged in illegal invoicing, email harvesting, and deliberate idle behavior.