
Citi's November online banking terms treat actions by third-party AI agents as customer-authorized, even after a credential leak. JPMorgan and Bank of America have tightened similar rules.
In brief
- Citi's November terms make customers responsible for what outside AI agents do with their banking logins, even if credentials leak.
- JPMorgan
JPM blocks outside AI agents from moving money, and CEO Jamie Dimon said there is a liability transfer. - Federal rules still limit customer liability for genuinely unauthorized electronic transfers, whatever a bank agreement says.
What happened
Citi updated its online banking terms in November so that actions taken by third-party AI agents count as authorized by the customer Citi's November agreement.
That holds even when an agent leaks the customer's credentials and someone else uses them to log in.
An AI agent is software that acts on a user's behalf, for example by navigating websites or starting payments, rather than only answering questions.
Citi's provision does not cover AI agents the bank offers itself, and Citi says it need not check whether an agent followed the customer's original instructions.
Other large banks moved the same way.
JPMorgan
JPM and Bank of America announced similar policies in October, and Citizens warned customers by email this week.
PNC changed its digital agreement on Wednesday, separating bank-approved connections from unrestricted account access.
JPMorgan
JPM blocks outside AI agents from moving money, and CEO Jamie Dimon said users face "a liability transfer", reportedly JPMorgan, Bank of America and Citi now warn YOU will be responsible.
The tightening follows an October 22 security scare.
Banks detected AI agents logging into customer accounts, sometimes taking steps that appeared to go beyond their users' instructions, and executives struggled to tell which agents were properly authorized.
Bank vendors have also been hit: a ransomware group leaked data on thousands of Visa {V} and Mastercard customers, and JPMorgan
JPM said attackers reached marketing data through outside providers a ransomware group leaked data.
Why it matters
The terms shift the cost of an agent's mistake from the bank to the account holder.
For anyone building agentic finance, that sets a precedent: an agent holding a login is treated as the user's own hand.
The same logic reaches crypto wallets, where a leaked key or a misdirected agent payment has no bank to absorb the loss.
Federal rules still require banks to investigate unauthorized electronic transfers and cap customer liability, and the Consumer Financial Protection Bureau says agreements cannot waive those protections.
How they apply to an agent the customer chose to connect is less clear.
What to watch
The Office of the Comptroller of the Currency has issued AI examiner guidance covering outside agents, payment controls and human approval.
The framework is nonbinding and creates no new requirements for banks.
Watch whether other banks copy Citi's wording, and how regulators treat losses when an agent exceeds its instructions.
Stablecoins such as USDC
USDC are a common settlement rail for agent payments, so the liability question follows the money off bank balance sheets.