Citibank changed its online user agreement this week to say that customers who hand their banking credentials to a third-party AI agent are responsible for what that agent does, including transactions it makes by mistake and losses that follow if the tool is hacked. A Citi spokesperson confirmed the change to TechCrunch and framed it as disclosure rather than a new restriction.
"As third-party AI tools become more widely available, we want customers to understand the risks associated with sharing banking credentials or authorizing those tools to access their accounts," the spokesperson said. The updated agreement, in the bank's words, reinforces that "customers who choose to grant account access to third-party AI agents are responsible for activities conducted through those agents."
Three more banks moved first
Citi is the third-largest commercial bank in the United States, and it is following rather than leading. JPMorgan and Bank of America, the country's two largest, announced similar policies last month. PNC, the eighth largest, changed its terms on Tuesday. Citizens sent customers an email this week with much the same message, according to a copy seen by TechCrunch.
What separates Citi's version is the hacking clause. The agreement does not stop at errors the agent makes on its own; it also places the loss on the customer if the AI tool leaks credentials or is breached, and someone else drains the account through it.
That is a notable place to draw a line, because liability for unauthorised transfers has generally sat with the bank, not the account holder. The new terms try to move the boundary by defining anything an authorised agent does as authorised activity. The wording does not change what happens when a password is stolen from a customer who never signed up for an agent, though banks have not published the operational detail behind the clause.
The competitive reason behind the paperwork
Banks have two motives here, and the second is easier to document. AI agents that read balances, move money and pick deposits compete with the products banks sell, and terms that make customers liable are one way to slow adoption without blocking it. Citi's agreement notes that agents the bank builds itself carry no such extra liability, which turns the clause into a preference for the in-house option.
The technical caution is not invented. Anthropic launched a personal finance tool earlier this month and billed it as "read-only," with no ability to move, manage or invest money, a limit that says as much about the state of the technology as about the vendor's caution. JPMorgan, meanwhile, has been testing agents that move money with no human involved, work that puts the same trust question inside the bank rather than outside it.
For customers, the practical result is a set of agreements that vary by institution at the exact moment agents become usable. A person who connects an assistant to a Citi account, a PNC account and a brokerage is now holding three different liability regimes, none of which is spelled out in the app where the connection is made. The banks have said what they will not cover. They have not said what they will.
The limited workaround is architectural rather than legal. Tools that can read an account but not move money sit outside most of the new language, which is why Anthropic's read-only design is a selling point and why the banks' own agents carry no extra liability. Anything with payment rights inherits the terms, and the terms now start from the assumption that the customer chose the risk.
The pattern extends past finance. Service providers across travel, retail and food delivery have spent the year adjusting their own terms as agents arrive on their side of the counter, and the payments infrastructure those agents depend on is the same one that seized up earlier this week.