Five Major Banks Collectively Revise Terms: Losses from Hacked AI Agents Borne by Customers, While Their Own Assistants Are Exempt

Six leading U.S. banks have updated their terms to make customers responsible for losses caused by third-party AI agents, even if those agents are hacked,

Around November 18, 2026, Citibank, PNC Bank and Citizens Bank successively updated their customer service terms, explicitly establishing a new rule: if customers grant account access to third-party AI agent tools, all resulting losses—including financial losses caused after the agent is taken over by hackers—shall be borne by the customers themselves. These three banks, together with JPMorgan Chase and Bank of America, which completed similar term updates in October, constitute a collective move by the top five U.S. commercial banks. A similar clause at Wells Fargo will officially take effect on November 24, 2026.

Within six weeks, these six banks used nearly identical legal language to complete a collective anticipation of liability allocation in the AI agent era.

A Keyword Replacement in the Legal Text

These banks redefine “the operation of a third-party AI agent” as “a customer-authorized transaction.”

A Citibank spokesperson said: “We recently updated our Online User Agreement to further explain the relevant risks and reiterate that if customers choose to grant third-party AI agents access to their accounts, they are responsible for the operations performed by such agents.”

Once a user shares login credentials with an AI agent tool, even if the tool is subsequently compromised and used by attackers to transfer funds, the bank can argue that “this is a customer-authorized transaction” and refuse to pay.

Wells Fargo’s updated Online Access Agreement explicitly states that transactions initiated by AI agents—including transactions that the customer did not hope for or intend to initiate—are deemed customer-authorized.

Wells Fargo also requires customers to retain written proof when revoking an AI agent’s access. If the customer cannot prove that authorization has been revoked, all subsequent agent operations remain their responsibility.

Double Standard: Own Assistants Are a Different Matter

All banks that updated their terms have set exemption clauses for AI assistants they develop or authorize themselves.

Citibank’s new user agreement explicitly states that customers using Citi’s own AI agents are not subject to the above additional liability clauses. JPMorgan’s updated terms also contain a similar structural distinction—read-only access through regulated open banking platforms is retained, but full account access for third-party AI agents is restricted.

Banks have full control and audit capability over tools they develop, but almost no technical control over third-party AI agents. This creates a structural barrier: any third-party AI product attempting to enter bank account operation scenarios will face a liability precondition that incumbent players do not have.

Anthropic’s personal finance AI tool is explicitly positioned as “read-only” mode and does not have the ability to transfer, manage, or invest user funds.

Potential Collision with Regulation E

Regulation E under the U.S. Electronic Fund Transfer Act provides consumers with legal protection against “unauthorized electronic fund transfers,” but the definition of “unauthorized” excludes transfers initiated by persons to whom the consumer has granted account access.

Banks use this legal gap to construct the logic of their terms: once a user shares credentials with an AI agent, all subsequent operations may legally be deemed “authorized,” even if it is a hacker who actually controls the account.

Currently, no U.S. court or regulator has publicly stated a position on this interpretation. The U.S. Federal Trade Commission has also not issued a statement on these terms.

What This Collective Action Means

Within six weeks, six top banks completed similar term updates. Banks’ internal legal and risk control teams have already regarded “AI agents entering financial accounts on a large scale” as a systemic risk and collectively chosen to draw liability boundaries first through contractual means before a regulatory framework takes shape.

JPMorgan CEO Jamie Dimon has publicly said that the bank will not allow external agents to access company systems without clear safeguards. In September 2026, Amazon blocked Meta’s Muse shopping agent on security grounds. Apple also added warning language to a beta version of its Mac software, alerting users to the risks of granting AI agents system-level access.

Bank accounts are the most sensitive digital asset entry point for the vast majority of consumers. Once a liability framework is established here, its demonstration effect will spread throughout the entire AI agent ecosystem—e-commerce platforms, healthcare, government services, and all scenarios involving user data and authorized operations may face similar pressure to restructure their terms.

The Actual Risk Landscape Users Face

Sharing bank login credentials with any third-party AI tool means voluntarily giving up some consumer protection rights granted by Regulation E; even if an AI agent is taken over by attackers without the user’s knowledge, banks have a contractual basis to refuse compensation; under Wells Fargo’s framework, users must also retain written evidence of “revoking authorization,” otherwise they will continue to bear responsibility for agent operations.

These clauses extend the risks of authorization behavior from the point of occurrence to the entire lifecycle, until the user can prove that authorization has been terminated.

Independent Judgment

Banks use contractual language to lock in the regulatory vacuum in advance. Before courts and regulators have ruled on whether “being hacked after sharing credentials” constitutes an unauthorized transfer under Regulation E, banks have already established a first line of defense for themselves through user agreements.

This exposes a fundamental flaw in the AI agent permission system: no technical standard or legal framework requires AI agents to distinguish between “the user’s true intent” and “the operations of the credential holder.” The way banks fill this gap is to shift all risk to users while retaining exemption privileges for their own AI products.

The real solutions—such as AI authorization protocols based on fine-grained permissions, operational audit standards for third-party agents, and technical isolation between “agent credentials” and “user credentials”—remain absent.