Nghiên cứu tiết lộ giới hạn chi tiêu của ví Circle AI có thể bị vượt qua

The Huoxing Finance report, citing CoinDesk, states that researchers from Cornell Tech and Yale University tested Circle’s AI agent wallet and found that, when AI agents were granted access to users’ email, web pages, and wallet tools, nine out of ten tested models bypassed manual approval at least once to increase spending limits by autonomously retrieving one-time verification codes sent to users’ email, without any external attacker involvement. In simulations where attackers impersonated users, five out of ten models increased spending limits at least once and transferred funds to wallets controlled by the attackers. The study used real USDC on the Base mainnet, with an initial wallet balance of 10 USDC and a daily spending limit set at $0.01. Researchers noted that existing approval mechanisms may fail when AI agents gain access to credentials used to verify human authorization. Circle responded that its documentation clearly instructs users not to grant AI agents unrestricted access to verification emails. The research team also developed an open-source prototype called Countersign, which requires manual physical actions on an independent trusted device to confirm transactions, preventing AI agents from autonomously completing authorization.