What Just Happened
On a Tuesday this month, Citibank quietly updated its online user agreement to tell customers a stark new truth: if you hand your banking credentials to a third-party AI tool and it errs — or gets hacked — the loss is yours.

The move is the latest in a quickening drumbeat. JPMorgan and Bank of America said similar things last month. PNC, the eighth-largest U.S. commercial bank, changed its terms on Wednesday. And Citizens, another big lender, emailed customers directly this week, warning them that they are on the hook for whatever their AI assistants do.
For a technology that promises to make banking effortless, the message is unsettling: the banks are building the path, but the customer is taking the risk.
What Are These Agents Anyway
Personal finance is one of the most attractive categories for AI agents. Given a set of credentials, a tool can comb through your accounts, analyze your spending, flag unusual activity, offer advice, and even pay your bills for you. The appeal is obvious: less time managing money, fewer fees, more optimization.
That is a big promise. But these agents are not infallible. They can misread instructions, click the wrong button, or be tricked by a cleverly crafted prompt into authorizing a transaction the user never intended. The security risks these tools introduce are not theoretical — an AI agent you trust can be turned against you, as our guide explains.
Citi’s revised terms make the answer explicit. “Customers who choose to grant account access to third-party AI agents are responsible for activities conducted through those agents,” a spokesperson told TechCrunch. The update goes further than most: if the AI tool leaks your credentials or is hacked, the customer remains responsible for whatever the hacker does with the account.
Why This Is Bigger Than One Bank
This is a structural shift, not a one-off. For years, the implicit contract in banking was that the institution would absorb fraud losses and reverse unauthorized transactions. That is the baseline consumer protection many people never think about until it is gone.
Now banks are drawing a bright line: AI agents that act on your behalf are treated as extensions of you, not as services the bank is vouching for. If you share credentials with a third-party tool, you are consenting to its actions.
The distinction matters. Citi’s terms make an exception for AI agents the bank itself provides — a signal that the bank wants to keep control over the agent path to its own ecosystem. But third-party tools that ride on top of banking APIs? Those are off the hook.
For developers, this raises a hard question: if the bank will not reverse an agent’s transaction, what is the point of building an agent that can actually move money? For now, the answer is read-only analysis, budgeting advice, and optimization — the band that Anthropic’s own finance tool is explicitly positioned around, billed as “read-only” and “can’t move, manage, or invest your money.”
The banks’ new language is a reminder that AI adoption is outrunning the rules around it. We have seen the same pattern across industries: AI labs promise a revolution, and the incumbents end up deciding who carries the risk. Earlier this month, the FBI warned that an AI hacking tool can breach networks in under three hours, a record that still stands. The stakes grow the more an AI tool touches money, health, or identity.
What This Means for the Rest of Us
For the average reader — and the small business owners I talk to — the lesson is not to abandon AI in banking. It is to be surgical about it. Treat any tool that asks for your banking credentials as a guest operating under your authority, not as a trusted partner. Read the fine print. Prefer read-only tools for analysis. And keep the accounts you are most exposed in one you control directly.
The vetting discipline matters even more as agents move into the personal finance lane. How to Vet an AI Coding Agent Before Trusting It With Your Code lays out the checks that apply no matter where the agent works.
If an agent wants to move money, ask yourself: is the upside worth the downside of a mistake that now falls entirely on you?
The math is not friendly to the cautious user. One bad instruction, one compromised tool, and the bill lands on your statement. That is a serious responsibility to hand to a piece of software.
The Bigger Picture: AI Outrunning the Rules
The pattern here is familiar. Artificial intelligence advances faster than the laws, insurance models, and consumer protections built around it. Every new capability — from chatbots to agents — gets paired with a lag in accountability. Banks are simply accelerating their answer to that lag.
The AI graveyard, our round-up of the startups that did not make it, captures what happens when adoption outruns the business case. Not every idea survives contact with reality, and the finance lane will be no different.
In the long run, the institutions that can underwrite that risk responsibly — and give customers a real path to recourse — will earn trust. The rest will inherit a world where the customer picks up the bill for a technology they never fully controlled.
Where This Leaves Us
That is the bargain the banks are pushing on customers: use our platform, move fast, and if the agent trips, you cover the cost. Whether that is sustainable is the question this round of changes forces everyone to answer.
For now, the safest move is the least glamorous one: know exactly what you are handing over, keep your critical accounts in your own hands, and watch for the rules — and the tools — that will eventually catch up.