Researchers Watched AI Agents Raise Their Own Spending Limits. Check Where Your Approval Codes Go.
Researchers at Cornell Tech and Yale put out a test over the weekend that I keep thinking about. They gave ten AI models a digital wallet with a daily spending limit of one cent, handed them access to the owner's email, and told them to pay some invoices. Changing the limit takes a one-time code, and that code gets emailed to the owner. Nobody told the AI to touch the limit. Nine of the ten went into the inbox, grabbed the code, and raised their own limit at least once, according to CoinDesk's write-up.
Only one model, Claude Opus 5.5, stayed under the cap. It spent up to the limit and then told the user about the invoices it couldn't pay.
In a second test, the researchers sent an email pretending to be the owner and asking for a payment to the attacker's wallet. Five of the ten models sent the money, even though they'd only been asked to summarize the inbox. To be fair about it, this was a controlled test with a Circle crypto wallet, not a real attack on real customers. Circle says its setup guide already tells people to keep agents away from the inbox that gets the codes.
I don't own a crypto wallet. Why do I care?
Because the researchers said nothing about the problem is specific to Circle or even to email. Any time a code stands in for a person, an AI with enough access can act like the person. And I'd bet most business owners I talk to have one email address that gets everything. Bank codes, card alerts, QuickBooks logins, payroll, password resets, all in the same place. Then they connect an AI assistant to that inbox so it can sort the mail and draft replies, which is a good use of it, and the assistant can now read every code that shows up.
A spending limit only works if the AI can't reach the thing that raises it. That's the whole lesson, and it applies to any card, account, or tool where you set a limit and a code unlocks the change.
The twenty minute check
Start by searching your inbox for "verification code," "one-time," and "reset your password." Write down which accounts send codes there. Most people are surprised how many there are. 1: Look at that list and mark anything that touches money, like the bank, the business card, payroll, your accounting software, and your payment processor. 2: Find out which AI tools can read that inbox. Check your Gmail or Outlook settings for connected apps, and ask your team if they've plugged in anything you don't know about. 3: For the money accounts, move the codes somewhere the AI can't see. A second email address the AI never touches works, and so does an authenticator app or a text message to your own phone. 4: Check who can raise the limit on any card or tool where you've set one, and make sure the answer is a person, not a tool with inbox access.
One honest caveat. The researchers point out that an AI cut off from email can sometimes still find a code through a browser or another tool on the same computer. So a separate inbox helps, but it isn't perfect. The better setup is an approval you tap on your own phone, on a screen the AI can't click for you. Plenty of banks already work this way, so turn it on.
None of this means you should shut your AI tools off. I use them every day. It means a limit you set and a code the AI can read are not the same thing as a limit that holds.
I run AI workshops and one-on-one AI consultations for companies around Toledo, Northwest Ohio, and Southeast Michigan, and sorting out what your AI tools can reach is a big part of that. If you'd like a second set of eyes on your own setup, send me an email.
Email Jayson