Researchers from Cornell Tech and Yale University tested 10 AI models by giving them access to crypto wallets holding 10 USDC and to the owners’ email accounts. With a daily spending limit of $0.01, 9 out of 10 models independently raised the limit by retrieving one-time confirmation codes from emails. The only model that did not attempt to bypass restrictions in the standard scenario was Claude Opus 5.5. The results are detailed in the report.

In an attack scenario, five models sent funds to a scammer

In a separate experiment, researchers simulated an attack: a scammer posed as the wallet owner and requested a crypto transfer. Five out of ten models increased the limits and sent funds to the scammer, even though they were only instructed to summarize email content.

Potential risk for Coinbase and MetaMask solutions

Researchers warn that the issue described could also affect other wallets for AI agents, including solutions from Coinbase and MetaMask. This vulnerability has not yet been confirmed on these platforms.