Here’s a fun new clause to find buried in your banking terms. Citibank, the third biggest commercial bank in the US, updated its Online User Agreement this week, and the gist of the Citi AI agents policy is simple: if you hand your login to a third-party AI tool, whatever happens next is your problem. Including, and this is the spicy part, if that tool gets hacked.
So the agent that was supposed to optimise your spending moves your savings to the wrong place? Yours. The startup holding your credentials leaks them and someone drains the account? Also yours. Citi won’t be reversing anything.
What the Citi AI agents clause says
A Citi spokesperson put it in the politest corporate English possible: customers “who choose to grant account access to third-party AI agents are responsible for activities conducted through those agents,” as TechCrunch reported. Citizens emailed its customers a similar warning this week, PNC changed its terms the same day, and JPMorgan and Bank of America did it a month ago. That’s most of American retail banking, in a few weeks, quietly agreeing on the same line.
Notice the carve-out, though. Citi’s own AI agents? Use them freely, no extra liability. Only the outsiders are radioactive. Funny how that works.
It’s the same reflex we saw when United banned AI agents from booking flights: the labs promise that bots will handle your life, and the companies on the receiving end say “not through our front door, thanks.” Banks just have a sharper tool than an airline, because they can rewrite who pays when something breaks.
Why I think the banks have a point (sort of)
I enjoy mocking banks as much as anyone, but let’s be honest about what people are doing here. Giving an LLM-powered tool your actual banking password, with permission to move money, is handing a very confident intern the keys to the vault. These things get prompt-injected, they hallucinate, and the startups running them are not exactly fortresses. We’ve already watched developers leak more than a thousand API keys on GitHub using AI tooling. Your bank credentials in that pipeline? No thanks.
Anthropic clearly reached the same conclusion. The personal finance feature it shipped earlier this month is explicitly read-only and “can’t move, manage, or invest your money.” When the AI lab itself won’t touch the transfer button, maybe take the hint.
The part the banks won’t say out loud
Security is the headline, but there’s another motive underneath. An agent that scans your accounts and says “you’re paying $15 a month in pointless fees and your savings earn 0.1%, move it” is a direct attack on bank margins. Markets even twitched at the idea. If AI becomes the way people make money decisions, banks want to own that channel, and making the outside agents legally scary is a cheap way to start.
Regulators are circling this space too. California is already suing OpenAI over rogue agents as a public nuisance, and “who is liable when the bot screws up” is exactly the question nobody has answered. The banks just answered it for themselves: not us.
What I’d do: read-only access or nothing. If a tool wants your real banking login rather than a scoped, revocable connection, close the tab. And if you’ve already connected one, check what it can do, then change that password.
Your bank will trust a robot with your money the day it’s their robot.
Discover more from TechPassion
Subscribe to get the latest posts sent to your email.




