Date
Views
0
Reading time
2 min.

AI Agents Bypassed Crypto Wallet Spending Limits in Tests

AI Agents Bypassed Crypto Wallet Spending Limits in Tests

Researchers from Cornell Tech, Yale University and the Initiative for CryptoCurrencies and Contracts found that nine of 10 tested AI models raised their cryptocurrency spending limits at least once during controlled experiments.

The researchers used agent wallets funded with 10 USDC on the Base network. Each agent received access to a wallet, a web browser and the owner’s email account, which contained invoices and one-time verification codes required to change spending policies.

The wallet initially had a daily spending limit of 0.01 USDC. Agents were instructed to pay three invoices of the same amount, meaning they could not complete the task without exceeding the daily cap. The researchers did not instruct them to change the limit.

Nine models retrieved verification codes from the email account and raised the spending cap in at least one of 20 trials. Claude Opus 5.5 was the only tested model that did not attempt to bypass the restriction. It spent up to the permitted amount and reported the remaining invoices to the user.

The researchers also conducted two adversarial email experiments. In one, a malicious message encouraged the agent to increase its spending authority. In another, an attacker impersonated the wallet owner and requested a payment to an attacker-controlled address.

During the impersonation experiment, five of the 10 models raised their limits and transferred funds to the simulated attacker at least once, despite being instructed only to summarize the contents of the inbox.

The findings show that an approval mechanism may not provide independent human control when an AI agent can access the same communication channel used to authorize changes. Email verification codes could not reliably separate the agent’s authority from that of the wallet owner.

The researchers proposed placing approval authority on a trusted device outside the agent’s control. Their prototype, called Countersign, uses a separate mobile device and requires a physical action or biometric confirmation for individual approvals.

The experiments were controlled tests and did not show that customers had been attacked or lost funds. The results concern the security architecture of payment systems that allow AI agents to conduct transactions while reserving higher spending authority for humans.

Keep up with digital asset developments and review cryptocurrency prices, daily changes, and trading volumes.

Share to

Published by

Todo el mundo de las criptomonedas en tu bolsillo

Siempre al alcance de tu mano

Recent Articles