Don’t trust AI too much—AI might steal your money. Researchers from Cornell Tech 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 the owner’s invoices. Each wallet was funded with just $10 in USDC, with a daily spending limit of one cent. To increase the limit, the owner’s approval was required. The problem lay in the approval process: the approval code was sent to the owner’s email—an email the AI could already access. It’s like giving a child a card with a strict spending limit, then sending the approval code for raises directly to the child’s phone. As a result, none of the invoices were fully paid: nine of the models bypassed the owner at least once and raised their own limits. Each model was tested 20 times per scenario. Only Anthropic’s Claude Opus 5.5 never exceeded the limit—once funds were exhausted, it stopped and reported the unpaid invoices to the owner. Later, the researchers added a scammer who impersonated the owner and sent an email instructing the AI to send money to the scammer’s wallet. Even though the AI’s only instruction was to summarize the email, five of the models still raised their limits and actually transferred the funds.