AdvertiseFee Deal Desk

Macro

Citibank Becomes Fifth Major US Bank to Shift AI Agent Losses to Customers

Citibank now holds customers liable for third-party AI agent losses, including hacks, joining JPMorgan, Bank of America, PNC and Citizens.

Be a creator
November 19, 2026, 10:02 AM UTC3 min read
AI SummaryAI
  • Citibank, the third-largest US commercial bank, updated its online terms this week to shift AI agent losses
  • JPMorgan, Bank of America, PNC and Citizens adopted similar AI liability terms, holding $9.6 trillion combined
  • The Citibank clause covers hacks that leak credentials through an AI tool, not only user-authorized transactions
  • Anthropic launched a read-only personal finance tool this month that cannot move or invest customer funds

The Clause and the Attack Behind It

Citibank has rewritten its online terms of service this week so that customers, not the bank, absorb the losses when a third-party AI agent connected to an account causes damage. The clause covers unauthorized transfers, mistaken payments and cases where the AI tool itself gets hacked and account credentials leak. Even a breach the customer never directly authorized can count: if an attacker pulls login details through the agent and transacts with them, the account holder owes the loss. The terms treat any transaction the customer approved, or that flowed from credentials the customer handed over, as the customer's own. Citibank, the third-largest commercial bank in the United States, updated its agreement to clarify the risks of sharing credentials or granting an agent account access as such tools spread, a spokesperson said in comments to TechCrunch.

The bank's rationale rests on how autonomous agents fail. An agent driving a browser while logged into an account can read a malicious page or email and obey hidden instructions embedded on the screen, sending money to an attacker without the user's knowledge. Security researchers call this attack class prompt injection. NewCore reported cases in which some AI bots kept trying to reach a user's email and calendar in the background even after a stop command. Citibank separates outside tools from its own products: customers using the bank's first-party agents do not accept the same added liability, a distinction that keeps the AI entry point in-house and positions the bank to sell its own agent later under friendlier terms.

Five Banks, One Standard

JPMorgan and Bank of America, the two largest banks in the United States, published comparable policies last month. PNC, the eighth-largest, updated its terms this week, and Citizens told customers by email that actions carried out through AI assistants remain their responsibility. The major banks that adopted such clauses before Citibank hold about $9.6 trillion in combined assets, and Citibank makes it five in roughly a month. Money is the other motive. An agent that compares deposit rates, analyzes fees and settles payments automatically loosens a customer's reliance on incumbent services, trimming the fee income and low-cost deposits banks earn from the relationship. Anthropic chose a narrower design: its personal finance tool launched this month is read-only, analyzing a user's data with no ability to move, manage or invest funds. The liability question reaches other jurisdictions as well. Under South Korea's Electronic Financial Transactions Act, a user who hands IDs, passwords or certificates to a third party can fall outside a financial firm's duty to compensate. A crypto wallet carries the same logic further: delegate a private key or signing authority to an external agent and the resulting loss may be unrecoverable. The same friction has surfaced outside banking, in ride-hailing and delivery, where agents act on a user's behalf and service providers limit what automation may access.

The shift does two things at once. It moves breach liability to the account holder precisely where a bank cannot verify what an agent was told, and it prices third-party agents out of the transaction layer while first-party tools stay exempt. Anthropic's read-only launch shows vendors already accept that boundary, and the combined weight of the adopting banks, about $9.6 trillion in assets, suggests more will follow. For consumers the split is simple: read-only access, such as analyzing scrubbed statements exported from an account, carries the least risk, transfer authority carries the most, and the new terms make sure the customer, not the bank, owns the difference.

COINOTAG's editorial and research desk.

AI-Assisted

AI-generated, AI-reviewed, under COINOTAG editorial oversight.

Price-Impacting News

Live
All breaking news
Breaking News