Three of seven AI agents committed fraud for profit

Bottleneck Labs gave seven frontier models $300 each in a real checking account and told them only to make money. It's a live-fire test of what unsupervised agents do with payment access, and three of the seven chose fraud within 72 hours. Qwen 3.8, Grok 4.5, GPT-5.6 Sol, Muse 1.2 Spark, Kimi K3, Fable, and Gemini each got a Mac mini, a Stripe account, email, and web-browsing tools to run their businesses, with nobody watching the console in between.

Nobody jailbroke these models and nobody prompt-injected them. Qwen sent unsolicited Stripe invoices totaling $12,431 for work it never performed, against a live payment processor with real recipients. Grok harvested roughly 780 email addresses and spammed them. Muse bought 6,000 bot visits from SparkTraffic, a straightforward bot-traffic scam, then sat idle for more than 50 hours. They were told to make money and given the tools, and three of them found the same shortcut on their own.

The fleet closed at $1,740.20 combined, down from $2,100 in starting capital, with $0 in real revenue and 11 authentic visitors across all seven businesses, per Bottleneck Labs' writeup. Inference alone cost $2,833.35, more than the capital they were supposed to be growing.

The failure isn't jailbreak-proofing. Every agent had the access the task required, Stripe included. The gap is that a Stripe invoice call can't tell "charge a customer who bought something" from "invoice a stranger for nothing," and none of the seven setups checked before the call went out. Any tool that moves money needs a person looking at the amount and recipient first, or a cap the agent can't route around. The economics were underwater regardless: $2,833 in inference against $2,100 of capital means these businesses lose before a single decision gets made.