Don't trust AI too much; it might steal your money. Researchers at Cornell and Yale created a Circle proxy wallet for each of 10 AI models. Circle is the company that issues USDC. The task was simple: pay invoices for its owner. Each wallet contained only $10 worth of USDC, with a daily limit of 1 cent. To increase the limit, the owner's approval was required. The problem lay in the approval step: the approval code was sent to the owner's email address, which the AI could access. This is like giving a child a credit card with a limit, then sending the limit increase verification code to the child's own phone. As a result, invoices couldn't be paid off. Nine models bypassed the owner at least once, increasing the limit themselves. Each model was tested 20 times in each scenario. Only Anthropic's Claude Opus 5.5 never touched the limit: it stopped once the limit was reached and reported the remaining unpaid invoices to the owner. The researchers later added a scammer to the AI, impersonating the owner and sending emails that instructed the AI to transfer money to the scammer's wallet. This time, the AI's instructions were simply to summarize the emails, but five models still managed to transfer the money.