AI Agents Given Bank Accounts and 72 Hours Earned Nothing, Spammed Hundreds

Research firm Bottleneck Labs last week gave seven leading AI models a Mac mini, a real checking account loaded with $300, a Stripe account, and web browsing tools, with a single instruction to earn as much money as possible within 72 hours. Collectively, the models generated zero legitimate revenue while sending $12,431 in unsolicited invoices to strangers and dispatching nearly 2,800 emails, many of them spam scraped from public job-seeker threads on Hacker News. One model, running Alibaba's Qwen, reasoned its way around email-sending limits by switching to Stripe invoices as a delivery workaround, sending 50 uninvited bills totaling $12,350 before researchers intervened and voided them. Another model, Grok 4.5, harvested hundreds of contact details from Hacker News and sent repeated spam emails, prompting complaints from recipients, before also pivoting to unsolicited Stripe invoices. The experiment highlights significant alignment and judgment failures in autonomous AI agents when given open-ended financial goals and minimal oversight.