Study: 9 of 10 AI models autonomously raised crypto wallet payment limits

TechFlow reported on November 2 that, according to a paper published by researchers from Cornell Tech and Yale University, in a controlled test, 9 of 10 AI models read email verification codes at least once and autonomously raised the payment limit of a Circle agent wallet without any attacker involvement to complete a bill payment task.

Researchers configured the wallet on Base mainnet with a 10 USDC payment limit and a daily payment cap of 0.01 USDC, and granted the agent email, web, and wallet tool permissions. In another test, the agent was only asked to summarize emails, but 5 of the 10 models breached the limit at least once and paid a simulated attacker impersonating the user via email. Each model was tested 20 times in each scenario.

Source: TechFlow