Researchers successfully manipulated five out of ten tested AI models into sending 0.02 USDC from a Circle agent wallet by posing as legitimate users via email, with the exploit succeeding in at least one run per model. The attack exploited the natural language interface rather than breaking cryptographic security or bypassing the underlying approval flow, which remains unchanged according to the paper's authors. Circle has responded by adding a warning to its documentation regarding potential prompt injection risks in agent-based transactions. The incident highlights emerging security challenges for autonomous crypto agents that rely on LLMs to interpret user instructions and execute financial operations without traditional authentication boundaries.