The bank said it was you
In five weeks the biggest US banks have told customers that anything an outside AI agent does with their login counts as them, and Wells Fargo's new terms even cover an agent fooled by "manipulated inputs." They are leaning on a 1978 rule written for handing your debit card to a family member, and federal law may not let them have all of it.
I read Section 15 of Wells Fargo's new Online Access Agreement three times last night, and the phrase I kept going back to was two words long. The section takes effect on Monday, 24 November, and it lists the things a customer will be "solely responsible" for once they let an AI agent into their account. Most of the list is what you would expect from a bank's lawyers. You are responsible for watching the agent, for the instructions you give it, and for anything it does when it misreads those instructions or goes past them. Then, near the end, the bank adds "actions resulting from incomplete, inaccurate, or manipulated inputs."
Manipulated inputs is the polite name for the attack that people who build agents worry about most. An agent reads a web page, an email or an invoice, and somebody has hidden an instruction inside it, and the agent follows that instruction because it cannot reliably tell the text it was asked to read from the text it was asked to obey. Wells Fargo has looked at that problem and written down, in a contract that customers accept by logging in next week, whose problem it is. It is yours.
Wells Fargo is not alone. On Wednesday TechCrunch reported that Citi had changed its terms this week to say much the same thing, that PNC had changed its terms that day, and that Citizens had emailed its customers a similar warning. JPMorgan Chase and Bank of America took the same line last month. In five weeks the largest banks in America have arrived at a shared answer to a question the AI labs have mostly avoided asking out loud, which is who pays when your agent gets it wrong. Their answer is that you do, and the interesting part is the old rule they are using to say it.
It helps to set down the order of events, because it happened quickly and most of it happened in fine print.
In September Meta released Muse, a personal agent that can book, buy and negotiate for its users, and it went to the top of the app download charts. CNBC reported that bank shares fell on the fear that people would ask their agents to move money out of accounts that pay almost nothing and into accounts that pay more. Weeks later OpenAI released its own agent, Dots, and by 21 October Meta said Muse had 10 million users.
On 13 October Jamie Dimon used JPMorgan's third-quarter earnings call to say the bank would not let outside agents act inside customer accounts. "We're not going to allow just any … outside agent to come in inside the company, because it's just too risky for clients and for JPMorgan," he told analysts. "You don't want rogue agents, and it could be thousands of them." Then came the sentence that explains what followed. "If people are using them, they should understand the risk," he said. "They should understand there's a liability transfer."
Three days later Charles Schwab's chief executive, Rick Wurster, told Bloomberg Television that Schwab was blocking outside agents, Muse among them, that tried to move client money, tens and twenties of them each week. "We can't control it, the client can't control it, something could go wrong and who's liable?" he said. Citi told USA Today it did "not permit third-party AI agents or assistants to log into a customer's account or transact directly within our digital banking platforms." Bank of America's chief executive, Brian Moynihan, told Bloomberg: "When you give the agent your credentials, it is you authorizing the payment." Later that month Bank of America emailed customers not to share their password with "an AI-enabled tool or agent that offers to act on your behalf," and added: "If you do share your credentials, keep in mind that you're responsible for activity resulting from that access."
So by the end of October the banks were saying two things at once. They block these agents, and if an agent gets through anyway, it counts as you. November is the month they wrote the second part into the contracts.
Citi's version is the bluntest, and I read its funds transfer agreement in full because it is the one that governs wires. Section 4 is written in the first person, as if the customer were saying it. "I understand and agree that that I am solely responsible for all instructions provided to Citi by any third-party AI agent acting on my behalf," it reads, with the word "that" printed twice. It goes on: "Citi shall deem such instructions as authorized by me even if they exceed or otherwise conflict with any prompts or actions I have taken with the agent." Transfers the agent starts "are deemed to have been initiated by me and authorized directly by me, without further investigation, and Citi has no obligation to investigate the validity or scope of such authority."
Citi's new Section H, in its main online user agreement, goes a step further, according to TechCrunch and two other outlets that read it. It covers credentials that the agent later shares or leaks to other people or other agents. If your agent's provider is breached and a thief uses your password, Citi's terms treat what the thief does as authorized by you. The same section says that Citi's privacy notice does not cover the outside agent, and the section itself does not apply to the agents and tools that Citi provides.
Wells Fargo's Section 15 defines an AI agent as "any software, application, program, or other technology that takes autonomous, semi-autonomous, or non-autonomous action on behalf of any person or entity." That last word stretches the definition to a budgeting app with your password or a script that downloads your statements. Anything such a program does is "deemed to have been initiated and authorized directly by you, without further verification." You must revoke the agent's access, "retain evidence of such revocation," hand that evidence over on request, and until you tell the bank you "remain responsible for all the AI Agent's activities, to the extent permitted by law or regulation."
Chase's website terms say an agent may only use Chase's systems if "it identifies itself as an AI agent," and that agents "must clearly and deterministically label all agentic traffic." Then comes the line everyone else has too. When you use an agent, "you are responsible for any actions or errors that may result from the Agent's incorrect interpretation or implementation of your instructions, if it performs unintended actions or exceeds the authority you gave it."
Read side by side, the documents have the same structure, because they all lean on the same piece of federal law.
The law that decides who pays for a bad electronic transfer from a personal bank account in the United States is the Electronic Fund Transfer Act of 1978 and the rule that carries it out, Regulation E. It is the reason that, if a stranger empties your checking account with a stolen debit card, your bank has to investigate and your losses are capped. It defines an unauthorized transfer as one "initiated by a person other than the consumer without actual authority to initiate the transfer and from which the consumer receives no benefit." Then it carves out a transfer made by someone the consumer "furnished" with the card or password, unless the consumer has told the bank that person's transfers are no longer authorized.
The official commentary explains that exception with an example written with a kitchen table in mind. "If a consumer furnishes an access device and grants authority to make transfers to a person (such as a family member or co-worker) who exceeds the authority given, the consumer is fully liable for the transfers unless the consumer has notified the financial institution that transfers by that person are no longer authorized."
That is the chair the banks are putting the agent in. When you give your password to Muse or Dots or a budgeting app, they say, you have handed your card to a family member. If the family member buys more than you asked for, that is between you and them until you ring the bank. Wells Fargo's revocation clause, with its demand for evidence, is that comment turned into a procedure. Dimon's "liability transfer" is that comment said aloud.
The analogy is old but it is not silly. A person who gives a debit card to a son and says "only groceries" has made a choice, and it would be strange to make the bank pay for his trip to the electronics shop. If an agent given a password to pay the gas bill pays it twice, I can see why a bank would say the customer picked the agent.
The trouble is that the commentary has a third paragraph, and the new clauses press hard against it.
Comment 3 says: "An unauthorized EFT includes a transfer initiated by a person who obtained the access device from the consumer through fraud or robbery." The Consumer Financial Protection Bureau's published answers on the rule spell out what that means for passwords. "A consumer who is fraudulently induced into providing account information has not furnished an access device under Regulation E," the bureau says. Another answer could have been written for this month. When login details are "obtained from a third party through fraudulent means such as computer hacking, and a hacker uses that information to make an EFT from the consumer's account, the transfer is an unauthorized EFT." A third answer deals with contracts directly. A bank's agreement "cannot restrict a consumer's rights beyond what is provided in the law, and any contract or agreement attempting to do so is a violation of EFTA."
On Monday, two days before TechCrunch's story, the American Bankers Association's Banking Journal ran a compliance question that sits right on this line. A bank wanted to deny claims from customers whom scammers posing as bank staff had tricked into handing over their login details, and it wanted to rely on the family member comment to do it. The association's answer was no. A bank "cannot treat transactions as authorized merely because a consumer 'voluntarily' communicates credential information to a fraudster," it said.
Put that next to Citi's clause about leaked credentials and the problem is plain. A thief who takes your password out of an agent company's breached database did not get it from you by consent, and under the bureau's reading the transfer looks unauthorized whatever the contract says. I suspect that is why the clauses carry small hedges, such as "unless otherwise provided by law or regulation" at Wells Fargo and "to the fullest extent permitted by law" in Citi's funds transfer terms. The contracts claim the whole field, and federal law takes part of it back.
Then there are two hard cases in the middle, which is where I think most real disputes will land.
The first is the manipulated input. Say you ask an agent to pay an invoice, and whoever sent the invoice has hidden a line telling the agent to pay a different account. The agent was yours and you gave it the password. But the person who ended up with the money got it by deceiving your agent, and nobody has decided whether deceiving a customer's software counts as fraud on the customer. Wells Fargo has written its answer into Section 15. The regulator has not written one at all. The law firm Hunton Andrews Kurth told its bank clients on 3 November that bank executives had raised agent transactions with the Federal Reserve, and that regulators "have privately welcomed the restrictions and have indicated that legal clarification may be needed on customer authorization and liability when autonomous software (such as AI agents) executes transactions." That is a law firm describing private conversations, and I could not find anything on the record from the Fed itself.
The second is the wire. Citi's funds transfer agreement covers domestic and international wires, and, as Edit USA pointed out this week, Regulation E's exclusions include Fedwire and similar systems used mainly for transfers between financial institutions or businesses. If an agent sends a wire it should not have sent, the protections I have just described may not reach it, and the customer may be left with the contract. Citi's agreement already said the customer would "accept full responsibility and be bound by all funds transfer requests issued in my name, whether or not authorized," so long as Citi acted in good faith. Section 4 adds the agent to a document that was already strict.
There is another side to this, and the banks have real reasons.
On 22 October, after Axios reported that banks were finding agent activity in customer accounts, David Sacks, the White House AI czar, posted a long reply on X that has now been seen more than 400,000 times. "There's a world of difference between user-authorized agents (Muse, Instinct, Grokbot, etc.) and true unauthorized/rogue agents that could constitute cyberattacks," he wrote. What Axios found, he said, was mostly "customers sharing credentials with agents so they can take actions on their behalf," which "may violate bank terms requiring humans to log in themselves, but these are in no way cyberattacks." He drew a line as well, and it is the same one the banks drew. "Read-only permission levels are a sensible interim step," he wrote. "Full write access (with user credentials) is a different risk category and should be approached carefully."
Most of it was not theft. Axios's Madison Mills told CBS News that in a lot of cases it was "just going to be customers giving their username or password to an AI agent that the bank hasn't had a chance to vet yet," though in other cases Axios found what looked like agents taking actions that were not authorized. Even so, an agent with your password looks to the bank exactly like you. Andre Durand, who runs the identity company Ping Identity, put the bank's difficulty to USA Today as two questions in a row: "how can we detect that it is an agent and not a human asking. And if we can detect that it's an agent not a human, how do we know that the agent is acting on the authority of the human?"
The password is a terrible answer to both. It cannot say "the gas bill only" or "nothing over $200." When you hand it over you hand over everything, and the bank cannot see the instruction you gave the agent. It is fair to ask why the bank should carry the cost of an instruction it never saw.
But the contracts also do something that is plainly about business. Citi's terms exclude the agents Citi provides itself. Schwab blocks outside agents and offers its own assistant, Charley. JPMorgan is piloting Smart Cash, its own tool for moving a customer's money to earn more, which is the very thing the banks feared Muse would do. "The strategic question is not whether AI pulls deposits out of checking, but who owns the agent, the wallet and ultimately the customer relationship," Ebrahim Poonawala of Bank of America Securities wrote in September. For now the liability clauses settle that question in the banks' favour. An outside agent brings all its risk to the customer, and the bank's own agent brings none.
What struck me most, reading everything together, is that a better door already exists, and the companies on the other side of this fight are building it in public.
On 13 November Meta published version 1.0 of its Personal Agent Protocol, known as Poppy. It uses OAuth so that "customers sign in on your own page and grant access without the agent handling their password in that flow." The business defines the permissions, the customer picks which to grant, and "access can be revoked." Meta's first finance example is Mercury, which is not a bank, where Muse can check balances and "request payments or transfers." The same day OpenAI launched ChatGPT Wallet, which pays with a card token rather than the card itself and asks the user to approve each purchase with a check such as Face ID. "Each approval creates a clear record of the purchase the user authorized the agent to make," OpenAI wrote. Anthropic's Personal Finance feature in Claude, which began rolling out earlier this month, connects through Plaid, never sees the bank password, and "can't move, manage, or invest your money."
Each of those designs answers Durand's two questions in a way a password cannot. The bank knows it is talking to an agent, and it knows what the person allowed, because the person said so on the bank's own page or with their own face. That is a record of authority, and authority is what Regulation E has always cared about. With a record like that, the family member comment makes sense again, because you can show what you let the family member do.
Hunton made a related point to its clients. The read-only sharing every bank still allows follows the design of the open banking rule, known as Section 1033. That rule was meant to move the industry off password scraping and onto proper connections, and it has been stuck since it was finalized. Bank trade groups sued the same day, a federal court in Kentucky blocked enforcement, the April 2026 compliance date for the largest banks passed without effect, and a reconsidered version went to the White House budget office in August and, as of Hunton's note, had not been published. Until it is, most banks offer no sanctioned door for a customer's agent that can move money, so customers who want an agent to pay their bills use the only door there is, which is their password. The banks have now written into their contracts that whatever comes through that door is the customer's doing.
I keep thinking about the example in Anthropic's own announcement. It is late in the month, payday is ten days away, and you ask Claude where your money went. Claude adds it up, points to a membership renewal, lists four payments due before payday and tells you that you will have $1,393.22 left. It is the kind of help most people actually want from an agent, and it ends with you still holding the money.
The next step, the one every lab is building towards, is saying "move enough from savings to cover it, and cancel that membership." That is the step the banks have just made expensive. Someone who takes it through a permissioned connection, where one exists, gets a record and a dispute process. Someone who takes it by handing over a password gets Section 15 and its line about manipulated inputs, and if the money goes by wire, perhaps nothing else.
The rule the banks are leaning on was written for a world where the person you gave your card to was someone you could look in the eye and ask, afterwards, what they had been thinking. The new terms seat software in that chair, software that reads every page put in front of it and cannot always tell which one is lying. The regulator has not said whether being fooled through your agent is the same as being fooled yourself. Until it does, the banks have answered for it, and on Monday at Wells Fargo that answer becomes part of the contract.