Executive Summary

Citi has added one clause to its customer terms. A customer who lets an outside AI assistant reach an account is treated as having authorized whatever that assistant does. If the credentials then travel further and land with another person or another assistant, the conclusion is unchanged. The clause went into Section H of the online terms of use revised in mid-November 2026.

Other banks have written the same sentence. PNC amended its terms on November 18, and Citizens sent customers a warning email that week. JPMorgan Chase and Bank of America announced similar policies a month earlier, in October. The five together hold $9.6 trillion in assets. The clause has one target, though. AI tools a bank builds and offers itself are left out of the added liability.

Sections 1 through 3 are what the reporting and the published contract language hold. Section 4 is this article's reading. The clause is more than a security measure. The bank has changed who is allowed into the account.

Key Figures

Three numbers carry the story. The size of the banks that moved, the loss cap the law sets for consumers, and one product design that avoids the problem entirely.

Source: TechCrunch reporting, the English edition of Digital Today, and secondary coverage carrying the contract language and law firm analysis.

$9.6 trillion

Combined assets of the five banks that amended terms or sent warnings

JPMorgan Chase and Bank of America gave notice in October, and Citi, PNC and Citizens followed in November

$50

Loss cap for a customer who reports a hack within two days

The protection set by the US electronic fund transfer rule. Report within 60 days and the cap rises to $500

Read-only

The other road Anthropic took

The Claude finance feature released on November 5 can read an account but cannot move the money in it

1

Handing Over the Credentials Covers Everything That Follows

Citi is the third largest bank in the United States by assets. In mid-November it added a new Section H, headed "Use of Third-Party Artificial Intelligence ('AI') Agents," to its online terms of use. The reach is narrow. It applies only where a third-party AI agent is used to access a Citi Online or Citi Mobile account or to use digital services. AI agents, tools and features that Citi provides itself are carved out in the text of the clause.

The clause opens with a warning. Do not allow a third-party AI agent to reach the account, do not give it information, and do not instruct it to carry out transactions. The sentence after that is the center of the provision. If a customer ignores the warning and hands online credentials to a third-party AI agent or allows it into the account, Citi treats every access and action by that agent as authorized by the customer.

The chain does not break there. According to reporting that carried the contract language, if those credentials are further shared or spread so that another person or another AI agent uses them, Citi still treats the access and actions taken with them as authorized by the customer. Handing the key over once covers what happens afterward. This is where the summary "you are on your own even if you get hacked" comes from.

The rest of the clause runs the same way. The duty to watch a third-party AI agent's activity sits with the customer rather than with Citi. If the agent misreads an instruction, does something it was never asked to do, or steps past the authority it was given, the customer carries it unless applicable law provides otherwise. A right pointing the other way went in as well. Citi may restrict, block or terminate a third-party AI agent's access without notice.

On the privacy side a gap opens. Allowing an outside agent in is itself deemed to authorize Citi to disclose customer information, sensitive information included, to that agent. And Citi's privacy notice does not apply to how that agent collects, stores, protects or shares the data. The clause gives its reason on the spot: this is territory Citi does not control.

A similar sentence appears to have gone in where money actually moves. One outlet reports that Citi also amended its online transfer agreement, so that a transfer an AI agent starts on the customer's behalf counts as authorized even where it departs from what the customer actually asked for. That passage rests on a single outlet. This article could not open Citi's own terms, and every quotation of the clause here leans on reporting that carried the language.

Read whole, the clause runs long, but liability splits at three points. When the customer first hands the credentials over, when the outside assistant uses them to enter the account, and when those credentials pass into another hand. All three are bundled as authorized by the customer, and only the tools Citi built stay outside the bundle.

Where the credentials travel Customer hands them over First delegation › Outside assistant acts Deemed authorized › Credentials leak onward Deemed authorized AI tools Citi built itself Outside Section H The split falls on access the bank controls and access it does not
▲ Pebblous original diagram. The reported Section H language laid over the path the credentials take
2

Five Banks Have Now Drawn the Same Line

This is not Citi alone. The amended terms PNC issued on November 18 are close to Citi's in structure. Using a third-party AI agent is the customer's responsibility, transactions the agent starts count as authorized, the bank has no duty to watch that activity, and access can be cut at any moment. Citizens chose email over terms. Its warning went to customers the same week.

Two banks had moved earlier. JPMorgan Chase and Bank of America gave notice of similar policies in October. The five did not line up and announce on one day. Two went first, and three followed about a month later.

By assets this is not a small group. These five are not, however, the five largest banks in the United States. PNC ranks eighth by assets and Citizens is smaller still. The summary that the five biggest US banks moved in the same direction does not match the asset ranking. What this is, precisely, is the first, second and third largest banks plus two mid-size ones.

Bank Assets When Form
JPMorgan Chase $4.09 trillion October Similar policy announced
Bank of America $2.65 trillion October Similar policy announced
Citi $1.98 trillion Third week of November Section H added to online terms
PNC $610 billion November 18 Terms amended
Citizens $233 billion Third week of November Warning email to customers

▲ Pebblous original table. Assets are the figures reported as of June 30, 2026, and the five together come to about $9.6 trillion.

It did not stop at five. Reporting says Wells Fargo amended its online access agreement effective November 24, and Charles Schwab has taken a similar direction. Both sit outside the five above.

The explanation the banks offer reads as a safety notice. As customers reach for outside AI tools more widely, the change is meant to tell them what risk they take on when they share credentials or allow account access. TechCrunch reporting carries a second motive alongside it. Banks are wary that AI tools could steer customers away from unnecessary fees or into higher-yielding products, which is to say that bank revenue could shrink. A security worry and a competitive worry sit inside one clause.

Another industry answered differently. United Airlines amended its contract of carriage on November 11 to bar unauthorized outside AI assistants from booking and ticketing altogether (related article). One side closed the door, and the other left it open while moving the liability to the customer. The banks took the second. The account stays open, and what changed is who receives the bill once something goes wrong.

3

Regulation E Has Its Own Definition of Authorized

Regulation E, the US rule on electronic fund transfers, defines an unauthorized transfer as one the consumer did not authorize and from which the consumer received no benefit. Law firm analysis reports the Consumer Financial Protection Bureau's position that a transfer made with credentials a third party stole falls inside that definition. Terms that say an action is deemed authorized do not carry the definition along with them.

The tension lawyers point to shrinks to one sentence. A successful login is not the same thing as approval of each transaction. The bank terms try to bind the two into a single act.

There is a trap on the other side as well. The same rule does not treat as unauthorized the acts of someone the customer handed the means of access to, until the customer tells the bank that the authority has been withdrawn. A customer who gave an AI assistant login details with their own hand may find that money the assistant sent to the wrong place is no unauthorized transfer at all, and hard to recover. This is the risk a consumer is more likely to meet in practice.

A real hack is a different story. Where a third party stole the credentials, the standard protections attach. On the reported thresholds, a report within two business days caps the loss at $50, and a report within 60 days caps it at $500. Law firm analysis reads the "on your own even if hacked" language as an attempt to contract around that safety net. One clause does not switch the rule off, but in a dispute the customer has to face the terms and the regulation at the same time.

What the law firms recommend to banks is procedure rather than wording. WilmerHale and Hunton Andrews Kurth advise reworking the dispute-handling process itself. An investigator should first establish when the customer handed credentials to the agent, and the scenarios involving fake agents or hijacked agents should be routed into the unauthorized-transaction process.

Moving the liability does not change the question that remains after something goes wrong. When did the customer delegate what, and did the agent stay inside that range? What answers it is the record, not the contract.

4

Why Pebblous Is Watching These Terms

Look again at the line Section H draws and one test is doing all the work. Can the bank control that access? The AI tools Citi built are tools whose permissions and behavior the bank holds. An outside assistant's actions and its data handling both sit beyond the bank's hands. So one stays outside the clause and the other becomes the customer's in full. Once agents started moving money, the first thing to settle was not a communication protocol or an authentication standard. The liability language in the contract arrived first.

Analysts read the motive differently. While outside assistants are pinned down by the clause, the entrance to the account stays on the bank's side, and room is left to attach the bank's own agent later on better terms. That reading appears nowhere in what the banks said, and it is an inference added by secondary analysis rather than a confirmed motive. Both explanations point at the same place, though. The authority to decide who gets to reach account data.

One design steps around the fight from the start. The Claude finance feature Anthropic released on November 5 reads accounts through Plaid and does nothing more. It cannot move money or invest it, passwords travel only through Plaid, and Claude never sees them. Keep write access out from the beginning and the question of who authorized what has nowhere to arise. Bank of America has warned customers not to share banking passwords with AI assistants, and this design settles on the product side how that warning gets kept.

The verification gap has already been said out loud on the supervisory side. Federal Reserve Governor Michael Barr, speaking in Milan on November 16, said large banks are reporting AI agents acting inside customer accounts, and that a bank cannot verify whether the agent is genuine, whether the customer consented, or whether the action matches the instruction the customer gave (related article). That came a few days before Citi and PNC amended their terms.

Between the two roads sits an empty field. Between an assistant that only reads and an assistant whose every act belongs to the customer lies the question of how the receiving side can check whose delegation this assistant carries and how far it reaches. The logins arriving at bank servers today leave that field blank. Whether a person typed it or a program did, and if a program, on whose behalf, does not travel with the request. With no way to check, a bank is left holding two handles. Open everything, or hand over all of the liability.

The gap between two designs Read-only Anthropic Claude Write + liability shift Citi, PNC, Citizens A design that verifies delegation Blank in today's logins
▲ Pebblous original diagram. Between read-only and full customer liability, a design that verifies delegation is missing

4.1Without a Record, the Scope of Authorization Cannot Be Explained

For the teams building and running agents, the immediate work lies in records rather than in reading clauses. When the delegation happened and how far it reached, how far credentials or tokens traveled, what the agent actually executed: all of it has to be kept in structured form from the start. The procedure the law firms recommend to banks assumes the same records. An investigator who wants to establish when a customer handed over credentials needs that moment written down somewhere.

Without records, a dispute ends as a fight over wording. The terms say the action is deemed authorized, and the customer says they never asked for it. With no evidence to separate the two, what remains is the sentence written by the side that drafted the contract. How to collect and organize data is an old problem. Keeping a record of who reached that data under what authority has only just begun to be asked for.

R

References

Primary Reporting

Legal and Industry Analysis

  • 3."Citi Makes Customers Responsible for Third-Party AI Agent Activity." unite.ai, Nov. 2026. Excerpts from the Section H text.
  • 4."Citi Puts AI-Agent Hack Losses on Customers; 5 Banks, $9.6T." FourWeekMBA, Nov. 2026. Per-bank asset figures.
  • 5."Major Banks Move to Block or Limit Customer-Directed AI Agents." Hunton Andrews Kurth, Nov. 2026. Regulation E legal analysis.
  • 6.Coverage of the "Claude Personal Finance" launch (online-tech-tips, 9to5mac, androidauthority, and others), Nov. 5, 2026.