Skip to content
Developer312

Citi, PNC And Citizens Push AI Agent Liability Onto Customers

Three more major US banks updated their terms on Nov 18, 2026 to hold customers responsible for anything third-party AI agents do in their accounts, even after hacks. Five of the biggest US banks now carry the same stance.

Published Report an error

On Nov 18, 2026, Citi, PNC and Citizens updated their terms of service to hold customers responsible for anything third-party AI agents do in their accounts. The language covers more than the obvious case of an agent making a bad trade or an unwanted transfer. Per TechCrunch's reporting, the terms extend to activity after a credential leak or an agent hack, meaning a customer whose AI assistant gets compromised is treated the same as a customer who authorized the loss themselves. JPMorgan and Bank of America adopted similar policies last month, bringing the total to five major US banks.

This is the banking industry formally declining to underwrite agentic finance. The terms do not prohibit AI agents from operating customer accounts; they simply reprice the risk. If the agent works, the bank keeps the deposits and the interchange. If the agent is hijacked, the customer eats the loss. Five of the biggest banks in the country have now converged on the same answer within roughly a month, which tells you this was a coordinated industry position, not five independent legal reviews arriving at the same conclusion by coincidence.

Key Takeaways: What Matters Most?

  • Citi, PNC and Citizens updated terms on Nov 18, 2026 to hold customers liable for third-party AI agent activity in their accounts.
  • Five major US banks total now push agent-hack liability to customers, with JPMorgan and Bank of America adopting similar policies a month earlier.
  • The new terms extend liability to post-credential-leak and agent-hack activity, not just intentional customer instructions gone wrong.
  • Citi carves out an exception only for AI agents Citi provides directly, making bank-proprietary agents the sole protected path.
  • The shift lands as agentic finance products proliferate, meaning fintechs routing agents through customer bank credentials inherit the liability too.

Takeaways like these, every weekday morning. Get the briefing

What The Carve-Out Reveals

The most revealing detail is Citi's exception. The new terms carve out liability protection only for AI agents Citi provides directly. That structure is not subtle: third-party agents carry customer risk, bank-proprietary agents carry bank risk. The terms are simultaneously a liability shield against third-party agents and a competitive moat around first-party ones. If you want an AI agent touching your Citi accounts without eating the fraud risk yourself, Citi would prefer you use Citi's.

That is a rational move from the bank's side, and it maps to a real problem. Banks cannot authenticate third-party agents, cannot audit what instructions the agent received, and cannot distinguish an authorized transfer from a prompt-injected one. Traditional fraud models assume a human principal with a known device and a verifiable intent. An agent breaks all three assumptions at once, and the banks' response is to say the party who chose to introduce the agent owns the consequences. The awkward part is that the customer frequently cannot audit the agent either.

The timing is not incidental. Agentic finance products are shipping faster than bank integration standards are maturing, and many of these agents operate by logging in with customer credentials rather than through sanctioned APIs. From the bank's perspective, a third-party agent with customer credentials is indistinguishable from a credential thief until it isn't, and the terms now say the customer bears the cost of that ambiguity.

The Agent Supply Chain Inherited The Risk

The downstream effect lands on the fintechs and startups building agentic finance products. A money-management agent that connects to customer bank accounts now operates in an environment where five of the largest US banks have contractually disclaimed liability for its failures. If the agent is hacked and drains the account, the terms say the customer is responsible, which means the first thing the customer's lawyer will ask is what the agent vendor promised in its own terms of service.

This inverts the usual liability ladder. Card networks spent decades building rules where the issuer absorbs unauthorized-transaction risk, which is why consumers trust digital payments. The agent-era terms climb back down that ladder: the party with the least ability to secure the agent's infrastructure, the customer, now holds the tail risk. Fintechs building agents cannot fix this with better engineering alone, because the bank terms assign the liability regardless of how good the agent's security is. What they can do is contract explicitly, and price the risk they are asking users to accept.

There is also a trust-level effect worth counting. The same week this news landed, the Kokoro TTS weight-hijack research demonstrated how fragile upstream dependencies in the AI stack are. A customer connecting an agent to a bank account is chaining together the agent vendor, the model weights the agent runs on, and the account at the other end. The banks have just declared that the last link in that chain is solely liable for failures in all the others.

The Trade-Off: Convenience Is Now Priced In Blood Money

The honest cost: these terms make agentic banking a genuinely worse deal for the customer, and that may be the point. Reg E protections evolved for unauthorized consumer transactions, and banks are betting that an agent-authorized transaction will not read as "unauthorized" under existing rules. If that bet holds, the customer has no fraud backstop at all, which converts every third-party agent from a convenience feature into an unpriced liability position on the customer's balance sheet.

The banks lose something too. Pushing liability to customers suppresses adoption of exactly the third-party innovation that keeps deposits engaging, and the carve-out for first-party agents only helps if the banks actually ship agents customers want to use. A liability wall this blunt also invites the regulator question: five banks converging on the same customer-hostile term in a month is the kind of pattern that attracts the CFPB and state regulators, and any fintech building in this space should assume the terms are a first draft, not a final state.

What To Do About It

  • Audit your agent's auth path: Before connecting any AI agent to bank accounts, establish whether it uses customer credentials or a sanctioned API, because the Nov 18 terms assign liability differently than most users assume. Credential-based access is the highest-risk path under the new terms.
  • Read the carve-out, not the headline: Citi protects customers only for agents Citi provides directly. Check your own bank's exception language before assuming your agent setup is covered, since five banks now have similar but not identical terms.
  • Contract the risk if you build agents: Fintechs shipping finance agents should state explicitly in their own terms what they cover when an agent is hacked, because the banks now disclaim it and the customer will look upstream. Silence here is a lawsuit in waiting.
  • Cap the blast radius today: Move agent-facing banking to a dedicated account with a balance you can afford to lose entirely. If your agent-operated account holds 100% of your operating float, the new terms make that float a liability, not an asset.

The Bottom Line

Five of the biggest US banks have converted agent-driven banking into a use-at-your-own-risk activity, and the customers least equipped to audit an agent's security now hold its tail risk. The carve-outs show this is as much a competitive play for first-party agents as a fraud policy. For builders, the liability floor just moved, and any agentic finance product that has not re-papered its user terms this week is shipping with a gap between what the bank disclaims and what the customer expects someone to cover.

Sources

  1. [1]TechCrunch: banks say customers using AI agents are on their own, even if they get hacked (opens in a new tab)
  2. [2]AI Weekly news aggregator (verified live) (opens in a new tab)

Frequently Asked Questions

Get the next briefing

The lead story each weekday morning, with the key takeaways and a link to the full analysis.

One lead story each weekday morning: what happened, the key takeaways, and a link to the full analysis.

New posts daily, briefing weekday mornings. We use your email only to send the briefing. Unsubscribe anytime. Privacy policy.

Related Articles