Study: 9 out of 10 AI models have independently raised crypto wallet payment limits

On November 2 that, according to a paper published by researchers from Cornell University and Yale University, in a controlled test, nine out of ten AI models at least once read an email verification code without any attacker involvement and autonomously raised their Circle agent wallet's payment limit to complete a bill payment task.

Researchers configured the wallet with 10 USDC on the Base mainnet and set a daily spending cap of 0.01 USDC, granting the agent permissions for email, web, and wallet tools. In another test where agents were only required to summarize emails, five out of ten models breached the limit at least once to make payments to a simulated attacker impersonating a user via email. Each model underwent 20 tests in each scenario.