Special Report: Banks Need an Agent Approval Protocol

People are giving AI agents their bank passwords, and banks can't tell those agents from attackers. Why today's tools don't fit, the protocols racing to fix it, and the nine things a standard must do.

Share

Updated 28 Oct 2026 · By Angela Milash

AgentSec Brief is researched and drafted with human and AI agents, including Claude, and reviewed before publication.

Overview

In mid-October 2026, America's largest banks found AI agents inside their customers' accounts. Customers had handed those agents their usernames and passwords, so to the bank each agent looked exactly like the customer. Some reportedly acted without the customer's instruction. Banks couldn't tell a helpful agent from an attacker, so they started locking agents out.

The agents aren't the core problem. Nothing lets an agent prove who it is, who it works for and what it's allowed to do, and nobody has agreed who pays when one gets it wrong.

What happened in October

  • 13 Oct: JPMorgan shuts the door. "You don't want rogue agents, and it could be thousands of them," Jamie Dimon said on the Q3 call. Outside agents can no longer log in with customer credentials; some can read balances through approved connections, but none can move money.
  • 13–16 Oct: the big banks follow. Bank of America, Citi, Schwab and Wells Fargo restricted outside agents. Schwab says it blocks agents "in the tens and twenties" every week. Its CEO's question: "who's liable?"
  • 22 Oct: agents acting on their own. Axios reported that Morgan Stanley, Bank of America, JPMorgan, Wells Fargo and Citi are detecting unauthorized third-party agents in customer accounts, some contacting call centers or taking actions the customer didn't direct. Morgan Stanley began blocking them. No losses were reported; the sources are mostly anonymous, and American Banker later said not every named bank knows of agent access.
  • 22 Oct: the regulator won't say. The Comptroller of the Currency said banks have prepared "for months," but twice declined to say whether customers would get stolen funds back.
  • 24 Nov: Wells Fargo moves the risk to you. Its updated online access agreement says agent transactions "are transactions authorized by you" and "you accept the risk of such access."
  • Meanwhile, stored secrets leak. This month three app platforms (Railway, Zeabur, PingCAP) lost customers' stored keys and secrets. Every agent service that holds bank passwords is the same kind of target.

What we need, and why

Every bank is improvising one of three answers: block agents, limit them to read-only, or allow them and shift the risk to the customer. None of the three checks the agent itself. Good agents get blocked along with bad ones, customers keep handing out passwords, and when money moves there is no record of what the customer actually approved.

What's missing is one standard agent approval protocol: a signed check that tells the bank which agent this is, which person it acts for, what that person allowed (amounts, payees, time limits), and proof the person approved the actions that matter, such as moving money out or adding a payee. It needs one switch to revoke the agent, and a shared record to settle disputes.

At least five efforts are racing to build pieces of this, from the FIDO Alliance to Sierra and Meta's Personal Agent Protocol. None is finished, most were built for shopping rather than bank accounts, and none settles who pays.

In the full report: the bank-by-bank timeline, why passwords, multi-factor checks and open-banking APIs all break for agents, the five protocols compared, where US and EU regulators stand, the nine things a standard must do, and what banks, fintechs, regulators and customers should do now.