Business•

Citi Rewrites Its Online Banking Terms So Customers Carry the Risk When an Outside AI Agent Misuses Their Login

Citi's November terms treat anything a third-party AI agent does with a customer's credentials as authorized, as PNC, Citizens, JPMorgan and Bank of America tighten their own rules.

Written By :MINRK
Citi Rewrites Its Online Banking Terms So Customers Carry the Risk When an Outside AI Agent Misuses Their Login

Photo by MChe Lee on Unsplash

1. Citi AI agents policy puts account holders on the hook

Citi has revised its online banking terms this November so that customers are responsible for whatever a third-party artificial intelligence agent does with their account credentials, including cases where the agent leaks a login that someone else then uses. The change became public on November 18, 2026, the same day PNC updated its own digital services agreement.

The revision lands at a moment when a practical question has moved from theory to contract language: if a piece of software is trusted to manage money and gets something wrong, who absorbs the loss? Citi's answer, at least where tools from outside the bank are concerned, is that the person who handed over the keys does.

The bank is one of several large US lenders that have spent the past month spelling out how they will treat software that logs in and acts on a customer's behalf. Taken together, the new policies amount to an industry-wide warning that sharing banking credentials with an autonomous tool is treated very differently from falling victim to a stranger who steals them.

2. What do Citi's new terms say about AI agents?

Citi's updated agreement says that any action carried out with credentials a customer gave to an outside AI agent counts as authorized by that customer. That classification holds even when the tool disregards the instructions it was given, goes beyond the permissions the customer intended to grant, or passes the credentials along to another party.

The terms go a step further on the question of intent. Citi states that it is under no obligation to look into whether an agent actually followed the prompts its user originally typed, a position the bank qualifies as being subject to applicable law. In practice, that means a dispute built on the argument that the assistant was told to do one thing and did another is not one the bank commits to examining under its own contract.

A Citi spokesperson described the November revision as a clarification of the risks that come with sharing banking credentials or giving outside tools access to an account. The spokesperson's framing presents the language less as a new stance than as a more explicit statement of an existing one: credentials are meant to be used by the account holder, and handing them to anyone or anything else carries consequences.

One boundary is worth noting. The provision is aimed at tools that come from outside the bank. It does not extend to AI agents that Citi itself offers to its customers, which remain governed separately.

3. Why are banks tightening their rules now?

The wave of contract changes traces back to a security scare on October 22, when major banks detected AI agents logging into customer accounts and, in some instances, taking actions that went beyond what the users had asked for. Executives said afterward that they had trouble working out which of those agents had been properly authorized and which had not.

That difficulty goes to the heart of how these tools differ from the chatbots most people are familiar with. A chatbot answers questions. An agent acts: it can move through websites, open financial records and start payments on a user's behalf without the user clicking each button. The convenience is obvious, and so is the exposure. A chatbot that states the wrong account balance has produced a mistake that can be checked and ignored. An agent that misreads an instruction, or whose login has been compromised, can move money.

From a bank's point of view, an agent that signs in with a customer's username and password looks, at the level of the login itself, much like the customer. The systems that decide whether to approve a transfer were built around the assumption that the person presenting valid credentials is the person entitled to use them. Software that holds those credentials and makes its own decisions about what to do next strains that assumption, and the October episode showed that the strain was no longer hypothetical.

The contract revisions are the banks' first formal response to that problem. They do not attempt to identify or block every agent. They settle in advance how responsibility is allocated when an agent that a customer chose to trust behaves in a way the customer did not expect, or exposes a login to a third party.

4. How JPMorgan, Bank of America, Citizens and PNC compare

Several of Citi's largest competitors have moved in the same direction, though not with identical wording. JPMorgan and Bank of America announced comparable policies in October. Citizens Bank emailed its customers this week to warn them that they could be held responsible for what their AI assistants do.

PNC changed its digital services agreement on Wednesday, November 18, and its approach is the most restrictive of the group. The bank draws a line between two ways an outside tool might reach an account. Connections that the bank has approved are permitted and provide limited access. Giving an agent unrestricted access to the account is a different matter.

Under PNC's rules, customers may use the approved connection methods but may not hand third-party AI agents their passwords or the one-time security codes used to confirm a login. Agents are also generally barred from initiating payments unless PNC has expressly allowed it. Where Citi's terms mainly address who is liable after something goes wrong, PNC's put more weight on restricting what an agent is allowed to do in the first place.

The differences matter for customers who use more than one bank. A setup that one institution tolerates, with liability attached, may be a breach of another institution's agreement outright.

5. Can a bank agreement override federal fraud protections?

No. Contract language of this kind does not automatically strip away the consumer fraud protections that federal law provides. Under federal consumer rules, banks generally have to investigate electronic transfers that a customer reports as unauthorized, and a customer's liability for fraud is limited. Those protections cannot simply be signed away in an account agreement.

The Consumer Financial Protection Bureau (CFPB), the federal agency that oversees consumer financial products, takes the position that banks may not use their agreements to waive these safeguards. What the bureau's rules do is draw a distinction between two situations. One is theft by an outsider who had no right to touch the account. The other is a transfer that the customer, or someone the customer gave access to, actually authorized.

That distinction is precisely where AI agents create uncertainty. An agent that a customer deliberately set up and supplied with credentials is not an outsider in the ordinary sense. How the federal protections apply to transactions carried out by such an agent is considerably less settled than how they apply to a conventional account takeover.

Whether a particular AI-related loss qualifies as unauthorized therefore turns on how the transaction came about. A payment an agent made while exceeding its instructions, and a payment made by a criminal who obtained a login the agent leaked, may be assessed differently even though the customer wanted neither. Citi's own qualifier, that its position applies subject to applicable law, leaves room for exactly that kind of case-by-case outcome.

6. Where regulators stand

Federal bank supervisors have begun to address the issue, though so far without imposing new obligations. Since the October incident, the Office of the Comptroller of the Currency (OCC), which charters and supervises national banks, has issued guidance for its examiners on artificial intelligence. The material covers outside agents, controls on payments and the role of human approval.

The framework is nonbinding and does not create new requirements for banks. It tells examiners what to look at, not what institutions must do. That leaves the banks' own contracts, for the moment, as the most concrete statement of how agent activity will be handled, alongside the long-standing federal consumer rules that those contracts cannot displace.

The result is a layered picture. Individual banks are defining agent activity as customer-authorized. A federal supervisor has signaled what it expects its examiners to review. And the consumer protection rules that predate autonomous software continue to apply, with open questions about how they map onto a transaction that a machine initiated on a person's behalf.

7. What the changes mean for customers

For account holders, the practical effect of the new terms is that the choice to connect an outside AI tool to a bank account now carries explicit contractual weight. At Citi, credentials shared with a third-party agent make the customer answerable for what follows, whether the tool misbehaves or the login ends up in someone else's hands. At PNC, sharing a password or a one-time code with such a tool is not permitted at all, and only bank-approved connections with limited access are allowed.

The policies also separate two categories of software that can look similar to a user. Agents offered by a bank itself sit outside Citi's provision. Agents from other companies, connected by handing over a login, sit squarely inside it.

None of this settles how an actual dispute would be resolved. The banks have stated their positions, the CFPB maintains that statutory protections cannot be waived by agreement, and the treatment of transactions made by agents that customers themselves authorized remains unclear. Until that question is tested, the safest reading of the new agreements is the literal one: a customer who gives an outside agent the means to log in should expect the bank to treat the agent's actions as the customer's own.

Related Articles

NEWSLETTERS

Don't miss another story.

Subscribe to the MINRK Newsletter today.

Tick the newsletters you want. No selection? You'll get MINRK Headlines.

By signing up, you will receive emails about MINRK products and you agree to our terms of use and privacy policy.

MINRK Daybook

Market analysis for crypto traders and investors.

EVERY WEEKDAY

MINRK for Advisors

Defining crypto, digital assets and the future of finance for financial advisors.

EVERY THURSDAY

MINRK Protocol

Exploring the tech behind crypto one block at a time.

WEEKLY

MINRK Long & Short

A must read for institutions. Insights, news and analysis delivered weekly.

EVERY WEDNESDAY

MINRK Headlines

The biggest crypto news and ideas of the day.

EVERY WEEKDAY

MINRK Policy

Examining the intersection of cryptocurrency and government.

WEEKLY

MINRK Research

Join thousands of readers who rely on MINRK for data-driven insights on the latest digital asset trends.

MONTHLY