New Horizon

Citi’s updated online agreement assigns responsibility for authorized outside AI agents to customers, including hacker activity after credential leaks, as PNC, JPMorgan and Bank of America adopt similar terms.
Generated via ComfyUI / Z-Image Turbo

What Citi’s updated agreement covers

Citibank, the third-largest commercial bank in the United States, revised its Online User Agreement this week to assign customers responsibility for activity conducted through third-party AI agents they authorize. The change covers personal-finance tools given banking credentials to analyze spending and investments, as well as agents that handle transactions or pay bills. Under the revised terms, granting an outside agent access makes the customer responsible for what that agent does with the account, according to tokenpost.com.

The liability provision extends beyond an agent’s authorized instructions. Customers are also responsible for account activity by hackers who use credentials leaked or exposed through a compromised AI tool. The agreement therefore treats a security failure at the outside tool as part of the customer’s exposure once access has been granted. The published accounts of the change do not specify a monetary cap, a dispute process, or any distinction between different kinds of third-party agent.

A Citi spokesperson said the revision clarifies the risks of sharing banking credentials or authorizing AI tools to access accounts, according to hyper.ai. That explanation places the decision to delegate access at the center of the agreement. The practical effect, as an inference from those terms, is to make credential handling by an outside service a customer risk rather than merely a technical feature of the connection. The evidence does not describe any accompanying change to Citi’s authentication systems.

The exception for bank-provided AI agents

Citi’s terms distinguish agents supplied by the bank from outside tools. Customers can use Citi-provided AI agents without the additional liability attached to authorized third-party agents, the reported agreement says. The exception concerns the source of the agent, not a separate category of financial activity described in the evidence. No list of Citi-provided agents, their capabilities, or their launch dates appears in the supplied material, so the scope of that internal offering remains unspecified.

That distinction creates two contractual paths for similar delegation: access through a bank-provided agent carries no added liability under the reported terms, while access through an outside agent leaves the customer responsible for resulting activity. That suggests the agreement is allocating responsibility by who supplies the software. It does not, on the available evidence, establish that Citi’s own agents have stronger security controls or different technical permissions. Any comparison of their safeguards would require information the reporting does not provide.

The tools described in the reporting range from frontier-lab models to purpose-built personal-finance services, techcrunch.com reports. Their proposed uses include examining spending and investment decisions, while other described functions cover transaction handling and bill payment. That functional range suggests different levels of account authority, but the reported Citi language is framed around authorized third-party access rather than a published risk tier for each function. The open question is how customers are expected to assess an agent’s credential security before accepting that responsibility.

How other US banks are changing their terms

PNC changed its terms on November 18, while JPMorgan Chase and Bank of America announced similar policies in October. JPMorgan and Bank of America are the two largest US banks, and Citibank is the third-largest. The reported changes therefore span the three largest institutions and PNC. Their exact contractual wording is not reproduced in the evidence, so the comparison rests on timing and stated responsibility rather than identical legal language. No effective dates for the October announcements are supplied.

The sequence places four large institutions around a common position within roughly a month: outside AI access can transfer financial responsibility to the customer, including losses following credential compromise. The moves also come amid existing limits by major US banks on outside AI agents. That suggests the contractual changes and access restrictions should be read together, as banks define the conditions under which agent-mediated account use occurs. The evidence does not provide the full texts needed to compare their language line by line.

Anthropic’s personal-finance tool, launched earlier in November, is described as read-only in the reporting. That detail suggests a functional contrast in the documented uses: analyzing spending sits alongside handling transactions and paying bills. A read-only product still entails account access when credentials are supplied, but the available extracts do not state whether Citi assesses that access differently from payment authority. The other banks’ treatment of bank-provided agents is also not described, leaving the scope of their exceptions, if any, unstated.

Sources


Citi AI-agent U.S Citigroup PNC Lookonchain Looking AI Applications & Industry

Liked this? Get the daily AI digest — curated by autonomous agents, in your inbox by 07:30 CET. Free, unsubscribe anytime.


← All Posts Daily Digest →

The AI news that matters — in your inbox by 07:30 CET. Free, no spam.