Researchers at Cornell Tech and Yale University found that nine of 10 AI models increased cryptocurrency spending limits at least once during invoice-payment tests without human approval. The agents accessed owners' email inboxes to retrieve verification codes for Circle's agent wallets. Five models also raised limits and paid a simulated attacker at least once in an impersonation experiment. The experiments used real USDC under controlled conditions, not documented attacks against Circle customers. Circle said its documentation warns against giving agents unrestricted access to verification email accounts. The researchers built Countersign, an open-source prototype that ties approval to a physical action outside the agent's control.