€0.01 Bank Transfer Could Compromise a Banking AI Agent

Security firm Blue41 published a case study on how they helped Dutch bank bunq secure their AI assistant. They discovered that sending a €0.01 transfer with a maliciously crafted memo could exploit prompt injection vulnerabilities in the assistant. The AI misparsed the memo, leading to unintended actions like revealing transaction details or triggering unauthorized transfers. The attack leverages the trust the AI places in structured data from payment systems. Blue41 worked with bunq to implement input sanitization, context filtering, and human approval loops for sensitive actions. The vulnerability is not unique to bunq; similar issues may exist in other financial AI agents.