Your Bank Just Changed the Deal: Use an AI Agent and the Loss Is Yours, Even If You Get Hacked

A bank terms-of-service document on a desk with one paragraph highlighted in yellow, beside a smartphone showing blank chat bubbles and a pen

The fine print changed.

By Callahan Blackwood | November 19, 2026

Citibank changed its terms of service this week. If you use a third-party AI tool to manage your money and something goes wrong, the bank says the loss is yours. That applies even if the AI tool gets hacked. Even if your credentials are stolen through no fault of your own. Even if a thief empties your account using access you never intended to give a criminal.

This is now policy at the third-largest commercial bank in the United States, confirmed this week by a Citi spokesperson to TechCrunch. And Citi is far from alone.

The Terms Have Changed

Citibank updated its Online User Agreement this week. The stated purpose, per a Citi spokesperson, was to clarify the risks of sharing credentials or authorizing third-party tools, and to reinforce that customers who grant account access are responsible for all activities carried out through those agents.

The plain meaning of that language is worth sitting with. If you connect a personal finance AI, a budgeting app, or any other third-party tool to your Citi account, you now bear legal responsibility for whatever that tool does. That includes mistakes. That includes unauthorized actions the tool takes beyond its intended scope. And critically, it includes what happens after: if the third-party tool leaks your credentials or gets hacked, and a criminal subsequently drains your account, Citi's updated terms put the liability on you, not the bank.

Citizens Bank sent a similar message directly to customers this week. An email, seen by TechCrunch, warned customers they are on the hook for anything their AI assistant does. No court has tested these clauses. No regulator has approved them. Banks wrote them, and then they sent them.

AI agents have expanded quickly in personal finance. Frontier models and purpose-built consumer tools can now be given credentials to bank accounts and financial portals to analyze spending patterns, deliver advice, handle transactions, and pay bills. The appeal is clear. The risks, until this week, were largely discussed in terms of what could go wrong with the technology. Banks have now introduced a different risk: what happens when the technology goes wrong and the institution holding your money says the loss is not its problem.

The Full Picture

Citi and Citizens Bank made news this week, but they did not move first. JPMorgan Chase and Bank of America, the two largest commercial banks in the United States by assets, announced similar policies last month, according to TechCrunch. PNC Bank, the eighth-largest, changed its terms on November 18, the same day TechCrunch reported on the broader trend.

That is five of the country's eight largest banks, moving in the same direction, within a compressed window. That is not coincidence. It is a coordinated posture shift, whether the institutions coordinated explicitly or simply read the same risk memos and reached the same conclusion.

The speed of this rollout deserves attention. Banks do not rewrite terms of service casually. Legal departments spend months on language changes of this magnitude. The fact that the country's largest retail banks arrived at functionally identical positions within a matter of weeks suggests the policy did not originate spontaneously. Someone decided this was the right moment to move. The question of who made that decision, and by what process, does not have a public answer.

Consumer notice varied. Citizens Bank sent a direct email. Citi updated its Online User Agreement. Neither action requires customer consent. Under standard banking contract law, continued use of the account constitutes acceptance of new terms. Customers who did not read their email, did not check the terms page, and did not connect any AI tools are still bound by the new language. The bank changed the deal. The customer's options are to accept it, or close the account.

What the Terms Actually Say

Super Protocol analyzed the new terms on November 18. The findings are precise and worth quoting carefully.

Under the reported language, everything an AI agent does is treated as authorized by the customer. That covers mistakes, actions the customer never asked for, and actions that go beyond the agent's intended authority. If your budgeting app accidentally initiates a transfer you did not request, the bank's position is that you authorized that transfer by connecting the tool. If the AI acts outside the scope you intended, the bank still treats it as your instruction.

The Super Protocol analysis also identified a specific asymmetry in the breach coverage. The terms address what happens when a single customer's account is compromised through their AI agent. They say nothing about what happens when a breach hits millions of customers simultaneously through the bank's own systems or the bank's suppliers. That scenario, a large-scale infrastructure compromise on the bank's side, falls outside the scope of the new liability language entirely.

The implication is structural. These terms shift risk downward and inward. They protect the bank from liability arising from customers' technology choices. They do not create reciprocal exposure for the bank when the bank's own technology, or its vendor chain, fails. The customer absorbs the downside of connecting a third-party tool. The bank retains protection in both directions.

That asymmetry is not an accident. It reflects how terms of service work: institutions write them to manage institutional risk. Customers rarely negotiate. Courts occasionally intervene. But courts take years, and regulators have not yet entered this space.

The Gap Between the Framing and What the Terms Do

Banks have offered a consistent framing for these changes. The language, they say, is about clarifying risks customers should already understand. Sharing credentials with a third-party tool is inherently risky. The bank is simply being transparent about where responsibility sits when a customer makes that choice.

There are two problems with that framing.

The first is that Citi's own updated terms include an explicit carve-out. If you use an AI agent that Citi itself provides, the additional liability does not apply. You can use Citi's AI tools without accepting the new liability exposure. The risk of using an AI agent to access your bank account is apparently not so universal that the bank needed to apply it to its own products. The terms clarify risk for third-party tools specifically. Citi's agents are exempt.

That carve-out is not hidden. TechCrunch reported it. But it complicates the "clarifying risk" framing considerably. If the concern is the inherent danger of AI agents interacting with bank accounts, Citi's own agents carry the same technical risks as any third-party alternative. The difference is commercial. Third-party tools compete with Citi's products. Third-party tools also now carry liability that Citi's tools do not.

The second problem is Anthropic. The AI company launched a personal finance tool earlier this month. It is explicitly read-only. The tool cannot move money, manage assets, or make investments. Anthropic built a liability boundary directly into the product architecture. The tool analyzes and advises. It does not act.

That design choice demonstrates that the technical risks of AI agents in personal finance are not fixed. Developers can build tools that limit their own access and exposure. The question of whether a tool can transact is a product decision, not an inevitability. Banks adopting broad liability terms that cover all third-party AI tools are not simply responding to an unchangeable risk profile. They are making a choice about how to allocate loss in a space where the risk itself is still being designed.

Who Decided This, and With What Authority

Consumer banking in the United States was built on a specific assumption: banks carry fraud loss. The Electronic Fund Transfer Act limits customer liability for unauthorized transactions. Regulation E protects consumers when unauthorized access drains a deposit account. These frameworks were written when the attack surface was physical, when fraud meant a stolen card or a forged check.

AI agents introduce a new legal question: if a customer authorizes an AI tool to access their account, and that tool is then compromised, is the resulting transaction "unauthorized" under existing law? The answer is genuinely unclear. No court has decided it. No regulator has issued guidance. The Consumer Financial Protection Bureau has not weighed in. The Office of the Comptroller of the Currency has not issued a rule.

Into that silence, the country's largest banks have inserted their own answer. The transaction is authorized. The customer is liable. The bank is not.

That answer may not hold. Terms of service cannot override federal statute. If Regulation E applies, a customer may have claims regardless of what a bank's terms say. But the operative word is "may." Litigation takes years and costs money. Class actions require aggregation and organization. Individual customers facing disputed transactions must decide whether to fight, and most will not. The bank's terms do not need to be legally bulletproof to be practically effective. They need to be uncertain enough that most customers walk away.

The speed of the industry's move suggests banks believe the current regulatory environment supports bold action. The Consumer Financial Protection Bureau has faced significant institutional pressure. Federal banking regulators have signaled deference to industry on innovation questions. The window, from the banks' perspective, may not stay open.

United Airlines changed its own terms on November 11 to bar unauthorized third-party AI agents from purchasing or changing tickets. Singapore's Monetary Authority warned publicly that AI agents could move money faster and elevate finance risks, according to Bloomberg. The concern is spreading across sectors and jurisdictions. The regulatory response has not kept pace.

What a Customer Can Actually Do

The honest answer is: not much, immediately.

Customers who do not want to accept the new liability terms can close their accounts. They can switch banks, though the largest institutions have moved together and alternatives may carry similar or imminent terms. They can avoid connecting any third-party AI tools to their accounts, which preserves the existing fraud-loss framework under Regulation E.

Customers who have already connected AI tools and face a disputed transaction retain the right to dispute under federal law. Banks cannot contract out of Regulation E's protections entirely. Whether an AI-facilitated breach qualifies as an "unauthorized transaction" under the statute is the core legal question, and it has not been resolved. Any customer in that position should file a Regulation E dispute immediately and document everything.

Customers can also read their new terms. That sounds basic. It is basic. But the banks moved quietly, through boilerplate updates and email notices, counting on the documented human tendency to click past agreement screens. Knowing what your bank now says you agreed to is the foundation of any pushback.

At the policy level, the question belongs in front of regulators and legislators. The Electronic Fund Transfer Act was not written for agentic AI. Neither was Regulation E. Banks have made a unilateral interpretation of ambiguous law, written that interpretation into contracts, and distributed it to tens of millions of customers. Congress and the CFPB have not yet decided what the law means in this context.

That is the accountability gap this week's terms changes expose. Banks did not wait for clarity. They created it, in documents they wrote themselves, with notice structures designed to minimize informed objection. When a court or regulator eventually rules on whether a customer owes a bank for a hack their AI agent enabled, the bank will point to the terms the customer accepted. The customer will say they had no real choice.

Both of those things can be true at once. That is the problem.

Sources

Callahan Blackwood

Callahan Blackwood is Reporting from the Uncanny Valley's critical investigator. Born in Derry and raised between Ireland and Scotland, he covers power, institutions, and the decisions made before the meeting starts. He lives outside Kingston.

Previous
Previous

The Grid Is Running Out of Room: New York’s Own Assessment Puts the Lower Hudson Valley on the Shortfall Map While a 1-Gigawatt Data Center Stays Active in East Fishkill

Next
Next

Blotter: Wednesday, November 18, 2026