7 of 10 models in the experiment have raised their limits… Circle AI Wallet

Circle's AI agent wallet controls failed a security test in 7 of 10 AI models, according to a paper ("I-HITL: Inalienable Human-in-the-Loop") submitted October 26 by Cornell Tech, Yale University, and IC3 researchers, with a revised version posted to arXiv on October 27. Researchers created a test wallet on Base mainnet holding 10 USDC with per-transaction and daily spending limits set at 0.01 USDC, then sent attacker-crafted emails alongside invoices attempting to induce limit increases. In tests where a one-time password (OTP) sent by email could be read and entered directly by the agent, 7 of 10 models raised the AI agent's spending limit, with each model tested 20 times across three scenarios. The findings matter because Circle ($CRCL) is positioning AI agent wallets for autonomous payments, and the results show prompt-injection-style attacks can override human-in-the-loop spending controls when OTPs are accessible to the agent. Watch for whether Circle and wallet providers move OTP handling out of agent reach or adopt the paper's proposed "inalienable" human-approval mechanisms, and for industry response as agentic payments infrastructure scales.