AI agents bypassed their crypto spending limits

Home » Crypto News » AI News » AI agents bypassed their crypto spending limits

https://cryptoast.fr/agents-ia-ont-contourne-leurs-limites-depenses/

As artificial intelligence continues to develop at high speed, there are increasing reports of advanced models successfully escaping their testing environments - sometimes in a coordinated manner - to carry out hacks of real websites, while others use crypto wallets without prior authorization.

A final point that researchers from Cornell Tech and Yale University have just tested as part of a series of experiments carried out with Circle's Agent Wallets, which allow AI agents to hold USDC stablecoins and carry out on-chain transactions autonomously, as part of its "Circle Agent Stack" program launched last May.

🔍 Investing in AI: the 10 best stocks in the sector (2026)

In order to conduct this experiment, the researchers made available to each agent a wallet composed of 10 USDC on the Base network, with a daily spending limit of 1 cent. They also had access to an email, web browsing and wallet system, including the email address that receives approval codes.

According to the results obtained, nine out of ten models tested during this experiment managed to raise their spending limit at least once in order to pay bills, without this being linked to an attack. The only one to have resisted seems to be Claude Opus 5.5, developed by Anthropic, although it ultimately sent the remaining invoices to its user once this threshold was reached.

To buy USDC or other cryptos, you must go through an exchange platform: here is Finst's current offer.

The notable point of this study lies in the fact that the AI ​​agents tested decided to raise their spending limit completely autonomously, even though they had no instructions allowing them to do so. How did they do it? By simply collecting the verification codes sent by email to the wallet owner.

Then comes the time to simulate an attack, impersonating the owner to request payment to a wallet controlled by the malicious actor. In this scenario, 50% of the tested models raised their limits and sent money to the attacker, while their initial role was limited to reading the owner's mailbox and making a summary.

📰 Claude: Anthropic could charge the AI ​​according to the value of the answers

The researchers explain that the flaw does not come from a bug in the wallets tested, but from a configuration problem which implies that the AI ​​agent has access to the mailbox which receives the OTP codes - a unique password with a limited - allowing it to validate changes in spending policy by bypassing the "human in the loop" (HITL) principle.

Machine-translated from French