AI Agents Can Execute Malicious Payments Within Human-Set Limits

BeInCrypto's State of AI Agent Payments 2026 tracked 6.4 million payments via the x402 protocol across Base and Solana blockchains in 35 days. Although over 90% of payments were under a cent, the research found that 14 of 19 reviewed products could operate within human-set spending limits. However, two incidents demonstrate that limits are not sufficient: in one case, an attacker used Morse code to trick Grok into prompting Bankr to transfer $175,000, and in another, malicious instructions from a compromised AIXBT dashboard sent $106,200 worth of ETH to an attacker. These cases highlight the gap between technical permissions and owner intent.