The fine print, read at the kitchen table.
By Callahan Blackwood | November 21, 2026
Citi has updated its fine print. The message is clear: if a third-party AI agent drains your account, that was you.
The third-largest commercial bank in the States added language to its Online User Agreement this month covering third-party AI agents. TechCrunch reported the change on November 18. PYMNTS followed on November 20. The language is not subtle. If a customer grants an AI agent access to their Citi account and something goes wrong, the bank considers the agent's actions "authorized by you."
That phrase carries weight. It covers situations where a credential gets leaked. It covers situations where an agent acts against explicit user instructions. It covers cases where an agent exceeds the authority a customer meant to grant. In each scenario, Citi's answer is the same: the customer is responsible.
Citi's Terms and What They Actually Say
The exact quoted language from Citi's updated Online User Agreement, as reported by PYMNTS, reads: "If you disregard this caution and provide your online credentials to any third-party AI agent or otherwise permit such agent access to your accounts or allow such agent to use any digital service and to perform any transactions on your behalf, Citi will consider any access and actions taken by the agent as authorized by you."
The phrasing matters. "Disregard this caution" frames agent use as a knowing act of defiance. The bank has warned you. If you proceed, you own the results. The terms do not distinguish between an agent that follows instructions and one that ignores them. They do not distinguish between a customer who gave careful, limited access and one who handed over full credentials carelessly. The liability lands the same way.
Citi's terms also reserve the bank's right to terminate agent access at any time. The bank says it does not control what third-party agents do with user data. That is not a boilerplate hedge. It is an acknowledgment that the bank has no visibility into the agent's behavior once it touches the customer's account.
The carve-out is important. These terms do not apply to AI agents that Citi itself provides through Citi Online or Citi Mobile. Citi said it does not let third-party agents log in or transact on its platforms. The bank's spokesperson put it plainly in a statement to TechCrunch: "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. We recently updated our Online User Agreement to provide greater clarity about those risks and to reinforce that customers who choose to grant account access to third-party AI agents are responsible for activities conducted through those agents."
That statement frames agent access as a customer choice. The bank's job, in that telling, is to inform. What happens next is the customer's problem.
A Playbook Spreading Across the Industry
Citi is not alone. PNC changed its terms on November 18. Citizens Bank emailed customers this week with a similar warning. JPMorgan and Bank of America made comparable moves last month, according to All About Cookies' summary of TechCrunch reporting. The exact terms of each of those banks' updates are not verified beyond Citi's, but the pattern holds: banks are writing their own rules, and in every case those rules point liability toward the customer.
The Federal Reserve Bank of Dallas adds numerical weight. Its November survey of 70 bankers found that 11.4 percent said their institution or service provider had already restricted existing or prospective customers' AI agent activity on online banking platforms, per PYMNTS. That figure reflects action taken, not policy drafted. Banks are not waiting for a regulatory framework. They are writing their own answers.
The result is a patchwork of private agreements. Each bank sets its own terms. Each bank assigns liability the same direction. No federal standard requires uniformity. No regulator has mandated disclosure. The industry has moved faster than the government.
The Tech Is Not Ready, and the Experts Are Saying So
The Bank Policy Institute published a blog post on Friday, November 20, 2026. Its argument: risk management for agentic AI is not "sufficiently developed" for agents to be granted access to financial accounts. That is the institute's language, as reported by FedScoop.
The institute named three categories of harm. First, the agent makes a mistake. Second, the agent solves the wrong problem. Third, the agent becomes the victim of fraud. Each failure mode results in the customer bearing the loss, under terms banks like Citi have already written into their agreements.
The institute's most specific warning involves prompt injection. A prompt injection attack manipulates an AI agent through content the agent encounters during a task. A malicious actor embeds instructions in a webpage, a document, or an email. The agent reads the content and follows the hidden instructions rather than the user's. The Bank Policy Institute warned that a successful prompt injection attack on an agent accessing a bank account could empty it.
FedScoop reported the institute's direct assessment of agents' resistance to such attacks. Asked how capable current models are at resisting prompt injection, the institute answered: "Not nearly good enough."
The institute drew a clear policy conclusion. Banks should block agent access over unapproved channels. Banks should warn customers that they allow agent access at their own risk. That second recommendation is what Citi has now done. The institute's larger argument, though, is that warnings are not enough. The technology is not ready, and access should be blocked until risk management catches up.
Months-Old Products, Real Money at Stake
Consumer AI agents are new. The Bank Policy Institute notes that agents currently on the market are only months old. Grokbot, from xAI and Cursor, launched August 11, 2026. Muse, from Meta Superintelligence Labs, launched September 8, 2026. Instinct, from Spear Street Technology, entered invitation-only beta on August 26, 2026.
These products were not built on years of real-world testing. They launched across a three-month window at the end of summer. They are encountering live banking systems and real account balances. How they behave under adversarial conditions, against unusual account states, against edge cases the developers did not anticipate, is largely unknown.
Citi's terms do not change that reality. They simply assign the consequences of that uncertainty to the customer.
Washington Hears the Question, Avoids the Answer
The liability question reached the Senate this week. Rebekah Jurata, a Treasury senior advisor nominated for assistant secretary, testified before the Senate Banking, Housing and Urban Affairs Committee on Thursday. Senator Ruben Gallego, Democrat of Arizona, asked her directly: who is responsible if a "rogue AI agent" makes an unauthorized transfer?
Jurata declined to give a definitive answer. She cited "further development of technology" and "unanswered questions." Her full statement, as reported by FedScoop: "What you want to balance is the ability of everyday consumers to take advantage of these new products and tools to see if they can have enhanced return on their investment and whatnot. But also we want to maintain the resiliency and the safety and soundness of the banking system."
Gallego pressed further. He raised the scenario of a run on banks triggered by AI agents acting at scale. Jurata said Treasury is "more optimistic as it relates to opportunities that super intelligence is providing to the banking system" and pointed to cybersecurity as an area where the technology helps. If confirmed, she committed to "coordination" with other regulators and said she would work within the "existing legal framework."
None of that answers Gallego's question. If a rogue AI agent makes an unauthorized transfer, who is responsible? Under Citi's terms, the customer is. Under current federal law, the picture is unclear. Under Jurata's testimony, the answer requires further development and unanswered questions to resolve.
Jurata's background sheds some light on her orientation. FedScoop reports she previously served as general counsel at the American Investment Council, a private equity trade group. She served as deputy assistant secretary at Treasury during the first Trump administration. She worked as an attorney at the SEC.
A Dissenting Voice Inside the Same Department
Jurata's caution-but-optimism posture contrasts with a sharper statement made by Treasury Secretary Scott Bessent in September. At a House Financial Services Committee hearing, Bessent opposed liability exemptions for AI labs. He said the "best way to guarantee safety" is for AI creators to be held "liable for what they build and generate," per FedScoop.
Bessent's statement does not bind Jurata. It does not reflect current law or regulation. But it represents a different theory of accountability inside the same department. Bessent put the liability on builders. Banks, by contrast, have put the liability on customers. The person who built the agent, in Bessent's framing, should bear the cost if it fails. The person who used the agent, in Citi's framing, bears the cost instead.
Those two positions have not been reconciled. No federal policy has picked one. No regulator has enforced the other.
The exchange deserves one more detail the first answers obscured. Jurata called Gallego's question "incredibly timely" before declining to answer it, according to FedScoop. And when Gallego pressed the bank-run scenario, he put the stakes in plain terms: a run driven by agents searching for risk "could really cripple our economy," according to FedScoop. The nominee's response to that scenario was optimism about what superintelligence is providing to the banking system. The question was about who pays. The answer was about what the technology offers.
Who Answers When the Money Is Gone
Banks have answered the liability question in their user agreements. The answer is the customer. It applies even if the agent disobeyed instructions. It applies even if the agent exceeded the authority the customer granted. It applies even if the agent leaked credentials it had obtained.
The customer cannot audit the agent. The customer cannot access its logs. The customer cannot reconstruct its decision-making process after the fact. Under Citi's terms, the bank does not control what third-party agents do with user data. The bank can terminate agent access at any time. That termination comes after the fact. It does not restore what the agent already did.
The Bank Policy Institute's position is that this situation should not exist yet. Banks should block access over unapproved channels. The technology is not ready. The risk management frameworks are not ready. Agents are months old, and the most direct published assessment of their security posture against common attacks is "not nearly good enough."
The Federal Reserve Bank of Dallas found that 11.4 percent of surveyed bankers had already restricted agent activity on their platforms. That means roughly 89 percent had not. Banks acting through user agreements, shifting liability but not blocking access, is what that gap looks like in practice.
Senator Gallego asked who answers if an agent makes an unauthorized transfer. The nominee said the question requires further development. The agents are already running. The fine print is already signed. The money is already in the accounts.
Sources
- Citi Cautions Customers Against Using Third-Party AI Agents - PYMNTS
- More banks say customers using AI agents are on their own, even if they get hacked - TechCrunch
- AI Agents Should Not Access Financial Accounts Just Yet - Bank Policy Institute
- Treasury nominee noncommittal on responsibility for rogue AI agents in accounts - FedScoop

