Institutions that chain pens to desks don’t like to live on the wild side. That’s why banks are gaming out anodd scenario: What if people let AI agents bank for them, expecting them to do the right things on their behalf, and they…don’t?
Bracing for impact
Axios reported on Thursday that major US banks have noticed an increase in unauthorized third-party AI agents accessing customer accounts. And while that sounds like a whole lot of hacking going on, generally speaking, the uptick was more a result of people giving bank account access to AI agents, like Meta’s Muse app or OpenAI’s Dots. But that still poses threats:
- Most banks’ rules require humans to log in to their own accounts, and those rules are being broken.
- AI agents can present security risks, especially if they’ve been granted access to transfer funds.
- In some cases, they may stray from human instructions, and banks can’t necessarily undo what an AI agent does. If an unauthorized purchase is made, it may not be clear who will be held legally liable.
Banks like JPMorgan Chase, Bank of America, Citi, and Morgan Stanley have taken steps to restrict third-party AI agents from accessing their systems, Axios reported. Some have moved toward granting agents read-only access, meaning humans need to actually approve transactions—a move recommended by White House AI czar David Sacks.
(Not) Rogue ones: Banks are quick to point out that these AI agent problems are very different from the AI agent problem that busted onto the scene this past summer, when a swarm of rogue bots broke out of their OpenAI sandbox and hacked into open- broke in
Breaking the bank:However, financial institutions are also preparing for the possibility that frontier AI systems could expose cybersecurity vulnerabilities that hackers exploit before they’re patched. Some hackers have already been able to infiltrate major international institutions using AI tools, prompting stress tests at places like the Bank of England.—BC
