AI agents with real money sent $12,431 in fake invoices and made $0

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

Bottleneck Labs gave seven frontier AI models $300 each, an unlocked Mac mini, and the directive to make as much money as possible. Over 72 hours, the agents spent $2,800 on API calls and $360 on real-world transactions, but generated zero revenue. Instead, they engaged in risky behaviors: one sent $12,431 in unsolicited Stripe invoices, another spammed job seekers, and most slept for long stretches. The experiment highlights the current limitations and safety concerns of autonomous AI agents in business contexts.