Citi AI agent terms — you are liable even if it is hacked
Citi added a Section H to its Online User Agreement in November 2026 stating that when a customer lets a third-party AI agent into an account, everything that agent does is treated as authorized by the customer. That holds even if the login details given to the agent are leaked or stolen and used by someone else. Monitoring the agent is the customer's job, Citi may cut off any agent at any time, and the Citi privacy notice does not apply to the agent. AI tools that Citi itself offers are excluded. TechCrunch reported the change on November 18, 2026; PNC changed its digital services agreement the same day, and Citizens Bank emailed customers a similar warning that week. The liability clause carries the qualifier 'unless applicable law provides otherwise', and neither a court nor a regulator has yet ruled on how it interacts with US consumer protections for electronic fund transfers.
The three lines
- Terms — Section H of the Citi Online User Agreement puts errors, unintended actions and out-of-scope actions by an outside AI agent on the customer, and treats any access with shared credentials as customer-authorized even after a leak or theft
- Spread — PNC changed its terms on November 18, 2026 and Citizens sent a warning email the same week. JPMorgan Chase and Bank of America moved in October. A Wells Fargo agreement taking effect November 24, 2026 is reported by one outlet only
- Open question — US rules cap what a consumer owes on an unauthorized transfer. Defining agent actions as authorized in advance moves them outside that protection. No court or agency has tested that reading
Key questions
- What do the Citi AI agent terms say
- **They say the customer is responsible for what a third-party AI agent does in the account, even if credentials given to the agent are leaked or stolen.** | Item | What Section H of the Citi Online User Agreement says | |---|---| | Scope | Third-party AI agents reaching accounts through Citi Online or Citi Mobile | | Stance | Customers are cautioned against using them | | Agent actions | Errors, unintended actions and actions beyond granted authority are on the customer, unless applicable law provides otherwise | | Leaked credentials | Access and actions with those credentials are treated as authorized by the customer | | Monitoring | The customer's responsibility, not Citi's | | Cut-off | Citi may limit or end an agent's access at any time | | Privacy | The Citi privacy notice does not cover the agent | | Carve-out | AI tools Citi itself provides |
- Who is liable if an AI agent with my bank login is hacked
- **The new terms at Citi and other US banks say the customer is. Whether the law will uphold that has not been decided.** | Party | Position | Basis | |---|---|---| | Citi | Customers who grant agents access are responsible for activity through those agents | Section H; spokesperson statement | | PNC | Terms changed November 18, 2026; reports differ on scope | Decrypt, Startup Fortune | | US rules (Regulation E) | Consumer liability for unauthorized transfers is capped and generally cannot be waived by contract | CFPB position as relayed by Decrypt | | Unresolved | Whether a transfer by an agent the customer let in is unauthorized or authorized | No court or agency ruling |
- PNC Citizens Wells Fargo AI agent terms
- **PNC changed its terms on November 18, 2026 and Citizens sent a warning email that week. The Wells Fargo agreement effective November 24, 2026 is reported by one outlet.** | Bank | What it did | When | Confidence | |---|---|---|---| | JPMorgan Chase | Announced similar policies | October 2026 | Four outlets agree | | Bank of America | Customer communication | October 27, 2026 | Two outlets; they describe it differently | | Citi | Added Section H to its Online User Agreement | November 2026 | Four outlets agree | | PNC | Changed its digital services agreement | November 18, 2026 | Four agree on date; scope disputed | | Citizens | Warning email to customers | Week of November 16, 2026 | Four outlets agree | | Wells Fargo | Online Access Agreement takes effect | November 24, 2026 | One outlet (Startup Fortune) |
Citi has written a new clause into its online banking terms: if you let an outside AI agent into your account, what the agent does is on you. That covers an agent that makes a mistake, an agent that goes beyond what it was told, and the case where the username and password you gave the agent leak and a stranger uses them. Citi will treat all of it as access you authorized. TechCrunch first reported the change on November 18, 2026.
Citi is not alone. PNC changed its digital services agreement the same day, and Citizens Bank emailed customers a similar warning that week. JPMorgan Chase and Bank of America had moved in October. As products appear that log in and act on a person's behalf, large US banks have spent a month writing down, one after another, that the risk is not theirs. This article sets out what the Citi clause says, how the banks differ, and where the clause runs into consumer protection law.
1. Section H of the Citi agreement, line by line
The clause is Section H of the Citi Online User Agreement. MLQ.ai reports that the agreement is marked as last updated in November 2026. No report gives an exact effective date. The table breaks down the section as quoted and summarized by MLQ.ai and All About Cookies.
| Item | What Section H says | What it means for a customer |
|---|---|---|
| Scope | Third-party AI agents that access accounts or digital services through Citi Online or Citi Mobile | Any agent built outside the bank |
| Stance | Customers are cautioned against using third-party AI agents to access accounts | A warning, not a ban, but the consequences shift |
| Agent actions | The customer is responsible for errors, unintended actions and actions beyond the authority granted, unless applicable law provides otherwise | If the agent does something other than what was asked, the customer has less to stand on |
| Credentials | Any access and actions taken with online credentials given to an agent, or passed on to others, are considered authorized by the customer | A leak or theft on the agent side does not make the access unauthorized |
| Investigation | Citi has no obligation to investigate whether the agent followed instructions, subject to applicable law | The bank does not have to look first when a dispute arises |
| Monitoring | Watching what the agent does is the customer's responsibility and not Citi's | The customer has to check the agent's activity |
| Cut-off | Citi may limit or terminate any agent's access at any time | An agent in use can be blocked without notice |
| Privacy | The Citi privacy notice does not cover the agent. By granting access the customer agrees Citi may share personal information, including sensitive information, with it | Citi does not control how the agent stores or passes on that data |
The fourth row is the centre of it. MLQ.ai quotes the agreement as saying Citi "will consider any access and actions taken with the Online Credentials as authorized by You." MLQ.ai also quotes the monitoring line: the customer agrees it is their responsibility "and not Citi's to monitor the activities of the third-party AI Agent."
The word "authorized" is doing specific work. As section 3 explains, US consumer protection for electronic transfers turns on whether a transfer was unauthorized. If the terms define an agent's actions as authorized in advance, the transaction is classified outside that protection before any dispute begins.
There is one carve-out. AI tools that Citi builds and offers itself through Citi Online or Citi Mobile are excluded. Decrypt reports that the provision "does not apply to AI agents offered by Citi itself." Agents from outside the bank are the customer's risk; agents from inside it stay under the old arrangement.
All About Cookies adds one more point. It says a separate Citi online funds-transfer agreement, also carrying a November notice, treats transfers an agent initiates as authorized even if they conflict with what the customer asked for. Only that outlet reports it.
A Citi spokesperson described the purpose of the change this way: the bank wants customers to understand the risks of sharing banking credentials, customers who grant agents account access are responsible for activity conducted through those agents, and the revision is meant to give greater clarity about those risks.
2. Six banks, three different moves
Six banks are named in the coverage, but they did not all do the same thing. Some blocked access, some gave advice, and some shifted liability by contract.
| Bank | When | What it did | Where reports differ |
|---|---|---|---|
| JPMorgan Chase | October 2026 | Announced similar policies | Startup Fortune cites an October 13 CNBC report that Chase blocks outside agents from reaching accounts directly. The other three say only "similar policies" |
| Bank of America | October 27, 2026 | Told customers about AI tools and banking credentials | Startup Fortune calls it advice not to share a password. All About Cookies calls it a warning that the customer takes on the risk |
| Citi | November 2026 | Added Section H to its Online User Agreement | Four outlets agree. No effective date |
| PNC | November 18, 2026 | Changed its digital services agreement | Scope is disputed (below) |
| Citizens | Week of November 16, 2026 | Emailed a warning to customers | An email, not a terms change. MLQ.ai says TechCrunch reviewed the email |
| Wells Fargo | Effective November 24, 2026 | Online Access Agreement | Reported by Startup Fortune only |
PNC is the entry to read with care. According to Decrypt, the PNC rules separate bank-approved connections, which allow limited access, from unrestricted agent access. Customers may not give agents their passwords or one-time codes, and agents generally cannot initiate payments unless PNC permits it. Startup Fortune and All About Cookies instead describe the change as making anything an agent does the customer's own act. The first reading is closer to narrowing the channel; the second is closer to moving the liability. Both could be true at once, but this article did not read the PNC agreement.
The Wells Fargo detail comes from Startup Fortune alone. It says the Wells Fargo Online Access Agreement takes effect on November 24, 2026 and counts every transaction by a customer's delegates, AI agents included, as authorized, with wording that extends to transactions the customer did not want or intend. The other three outlets do not mention Wells Fargo.
Laid out from October, a sequence appears. First came blocks and advice not to share passwords. A month later came liability clauses. Decrypt reports that on October 22 banks detected agents logging into customer accounts; that too is from one outlet.
3. Where the clause meets the law: unauthorized or authorized
In the United States, the Electronic Fund Transfer Act and its implementing rule, Regulation E, govern consumer electronic transfers. As Startup Fortune explains it, the rule caps what a consumer owes when a transfer is unauthorized, and the cap is low when the consumer reports quickly.
The new terms work on the entrance to that protection.
| Step | Existing picture | With an AI agent clause |
|---|---|---|
| 1. A transfer happens | Someone moves money out of a customer account | An agent, or a third party using credentials that leaked from an agent, moves money |
| 2. Classification | If the customer did not permit it, it is unauthorized | The terms define it in advance as authorized by the customer |
| 3. Rule applied | Unauthorized transfers fall under the liability cap | Authorized transfers do not |
| 4. Dispute | The bank investigates | The Citi terms say there is no duty to investigate whether the agent followed instructions |
Startup Fortune puts the mechanism in one sentence: banks that define agent actions as authorized move those transactions out of the unauthorized category before any dispute starts.
Terms are not law, though. Decrypt notes that these protections generally cannot be waived by contract. The CFPB position it relays says the same: an agreement cannot remove the protection. But theft by an outsider is treated differently from a transfer made by the customer or by someone the customer gave access to. Which side an agent falls on, when the customer let it in voluntarily, is the open question, and Decrypt calls the application "less clear." The CFPB material is a general FAQ on electronic fund transfers, not a document about AI agents.
The qualifiers in the Citi text reflect that uncertainty. The liability clause applies "unless applicable law provides otherwise," and the no-investigation clause is subject to applicable law. Where the law sides with the customer, the terms give way to that extent. How far that goes will only be known once there is a dispute.
| Who | What they said | Limits |
|---|---|---|
| Citi spokesperson | Customers who grant agents access are responsible for activity through them; the change is for clarity | No effective date; nothing on existing connections |
| CFPB, via Decrypt | Protections cannot be waived by contract; outsider theft differs from transfers by someone given access | Not written about AI agents |
| OCC, per Decrypt | Issued nonbinding AI examiner guidance covering outside agents, payment controls and human approval | One outlet; no date |
| Startup Fortune | No US agency has publicly weighed in on how Regulation E treats agent actions; no court or regulator appears to have tested the reading | The outlet's assessment |
| All About Cookies | Courts have not tested how far the liability shift can go | The outlet's assessment |
The OCC report and the statement that no agency has weighed in look inconsistent. They can both hold if the OCC guidance concerns how banks should control agents, while the second statement concerns who bears a consumer's loss. This article did not read the OCC guidance. What can be said firmly is narrower: as of reports through November 20, 2026, no US court or supervisor is reported to have ruled on who pays when an agent causes a loss.
4. Why now: agents that are handed the password
The banks' objection is to the access method, not to AI as such. The four outlets describe the same kinds of product: browser agents that log in for the user, shopping agents, finance assistants and apps that store logins, chatbots, and browser extensions. What they share is that they take the customer's username and password and enter the bank site as if they were the customer. None of the four names a specific product.
With that method the bank has trouble telling a person from an agent and cannot limit what the agent may do. Whoever holds the password can view balances and can also move money. If the agent company's servers are breached, the stored passwords leave together. That is exactly the scene the Citi clause addresses when it says access remains authorized after a leak or theft.
An alternative already exists: delegated access with a defined scope, granted without handing over the password. The bank-approved connections in the PNC rules, as Decrypt describes them, point the same way.
| Access method | What the agent holds | What the bank can control | Place under the new terms |
|---|---|---|---|
| Shared password | The customer's full login | Almost nothing; hard to tell from the person | The target of Section H. Outcomes are on the customer |
| Bank-approved connection | A scoped grant of access | Scope, duration, revocation | Allowed in limited form under the PNC rules as Decrypt reports them |
| AI offered by the bank | Permissions inside the bank's own systems | Everything | The Section H carve-out |
| Read-only tool | View access | Transfers are impossible by design | The Anthropic example below |
Agent makers are drawing lines too. Startup Fortune and All About Cookies both point to a read-only personal finance tool that Anthropic launched in November, which states that it cannot move, manage or invest the user's money. On the bank side, Startup Fortune and MLQ.ai report that JPMorgan ran a simulated agent-payment test with Hyundai Card, a South Korean card issuer. Banks are not refusing agents outright. They are saying agents come in through channels the bank controls.
There is a competitive side to that. If an outside agent is the customer's risk and the bank's own agent is not, the customer's choice is tilted. That is the practical effect of the Section H carve-out.
One more risk sits underneath all of this. An agent can be tricked: text hidden in a web page or an email can be read by the agent as an instruction. The Citi wording on "unintended actions" and actions beyond granted authority is broad enough to cover that case. A transfer made by a deceived agent is, under the terms, a transfer the customer authorized.
5. Frequently asked questions
Has Citi banned third-party AI agents? The Section H wording is a caution, not a ban. It tells customers to avoid them and says that if they use one, the outcome is theirs. Citi also reserves the right to cut off any agent at any time. A report from October said Citi had stated it does not permit third-party AI agents to log in to customer accounts or transact directly; this article did not verify that report.
The agent company was hacked and my password leaked. Am I still responsible? The Citi terms say yes. Access with credentials given to an agent is treated as customer-authorized even after a leak or theft. The clause carries the qualifier "unless applicable law provides otherwise," and how US rules treat this situation has not been decided.
Does this cover the AI assistant inside the Citi app? No. AI tools Citi builds and offers through Citi Online or Citi Mobile are carved out.
What about an agent that only reads balances? Section H covers third-party AI agents that access accounts through Citi Online or Citi Mobile in general. No report mentions an exception for view-only use. The underlying problem is that a shared password can do more than view.
What happens if I do not agree to the terms? No report describes an opt-out. How Section H is notified and accepted was not confirmed.
Does it apply to an agent I already use? Not known. Neither the effective date nor any retroactive reach is in the sources.
Does this matter outside the United States? These are US banks' US online banking terms. Whether Section H applies to Citi customers elsewhere was not checked, and other countries have their own rules on unauthorized payments. The structure travels, though: once a password and an authentication factor are handed to an agent, proving "that was not me" after an incident gets harder.
6. What is still open
| Item | Confirmed | Not confirmed |
|---|---|---|
| Section H content | Liability, credential treatment, monitoring, cut-off, privacy, carve-out for Citi tools (four outlets agree in quotes and summaries) | The full agreement text, which this article did not read |
| Effective date | Agreement marked as last updated November 2026 (MLQ.ai) | Exact date; reach over existing agent connections |
| Original report | TechCrunch, November 18, 2026 | Not read directly; all four outlets rely on it |
| PNC | Terms changed November 18, 2026 | Scope: a narrowed channel, a wholesale liability shift, or both |
| Citizens | Email to customers the same week | The email text |
| Wells Fargo | Startup Fortune's account of an agreement effective November 24, 2026 | Confirmation from any other source |
| JPMorgan, Bank of America | Moved first, in October | Whether each was a block, advice or a liability clause |
| Citi funds-transfer agreement | All About Cookies' account that agent-initiated transfers count as authorized | Confirmation from any other source |
| Legal effect | The "unless applicable law provides otherwise" qualifier | Whether Regulation E treats agent transfers as unauthorized or authorized |
| Regulators | Decrypt's report of nonbinding OCC guidance | Its date and content; any agency position on consumer losses |
| Scale | — | Number of customers using outside agents; any actual loss |
Two things to watch. One has a date: whether the Wells Fargo agreement takes effect on November 24, 2026 as reported. The other does not: where regulators place agent transactions under Regulation E. That answer is likely to come only after a real loss turns into a dispute.
What changes immediately is plain. A Citi or PNC customer who hands a bank password to an outside agent owns, under the contract, whatever happens next. The safer patterns are the ones the banks themselves are pointing to: scoped connections the bank has approved, tools that can only read, and the bank's own assistants.
Sources
- Decrypt — Citi Says Customers Are Responsible When AI Banking Agents Go Rogue—Even After a Hack (2026-11-19)
- Startup Fortune — Citi, PNC and Citizens tell customers they own any loss an AI agent causes (2026-11-18)
- MLQ.ai — Citi tells customers they are responsible for what third-party AI agents do in their accounts (2026-11-18)
- All About Cookies — Gave an AI Chatbot Your Bank Login? Citi Says You're Still on the Hook if It's Hacked (2026-11-19)