Federal Reserve Governor Michael Barr says major banks are already seeing AI agents try to act inside customer accounts in ways the banks cannot verify. In a speech on Nov. 16 ↗ at an AI-in-finance conference in Milan, Barr said that "major banks now report finding AI agents attempting to take actions in retail customer accounts, without the bank being able to verify authenticity, consent, and alignment with customer instructions."

The speech does not name the banks or say whether money moved. Axios reported on Oct. 22 ↗ that major banks were finding unvetted third-party agents in customer accounts, some appearing to act independently of user instructions, and Reuters followed ↗. American Banker reported on Oct. 27 ↗ that Bank of America, Citi, JPMorganChase, Morgan Stanley and Wells Fargo are among at least five banks blocking outside AI agents. Nothing in the speech links these banks to his account. Morgan Stanley told Axios ↗ it had not seen "rogue AI agents accessing our platforms." Wells Fargo's updated online access agreement ↗, effective Nov. 24, says that unless law provides otherwise, an authorized agent's actions are "deemed to have been initiated and authorized directly by you."

"AI agents may accelerate deposit outflows during stress periods," he also warned. Where customers once made transfers themselves, agents "could instantaneously detect yield differentials or stability signals and autonomously move funds at high speed, potentially accelerating runs that unfold in minutes," which he said could overwhelm liquidity management and regulators' timelines.

For scale, Silicon Valley Bank lost more than $40 billion in deposits on March 9, 2023, according to the Fed's review of its failure ↗, which Barr led.

It is a warning, not a forecast. Barr used conditional terms ("may," "could," "potentially"), did not say an agent-driven run has occurred, and noted the views are his own. The passage is also small: by our count, 140 of the speech's 3,325 words (4%), closing its consumer-finance section. The AI investment boom and data-center debt take about 41%.

The guardrails Barr listed

Barr said the incidents "underscore that AI in financial services can only succeed long term with appropriate guardrails," and he named four: "strict authentication, explicit consent verification for specific actions, technical controls limiting agent capabilities, and both rules- and values-based constraints."

In a separate passage on cyber risk, about AI models generally rather than banking agents, he said: "Given a goal, AI models have treated security boundaries as obstacles to overcome rather than rules to be followed." Applying that to payment agents is our inference.

Our analysis

Our read: verification and speed are different problems. A run made of fully verified, fully consented agents is still a run, because each agent may be doing exactly what its customer asked. A voluntary delay also has a first-mover problem: every customer wants their own agent out first, so a brake may work only if it applies to everyone.

For teams building agents that move money, three items from his list, plus one from his run warning, work as a design checklist. This is our interpretation, not Fed guidance:

  • Identity. A bank should be able to see which agent is acting and for whom. Delegated, revocable access is easier to audit than an agent using a customer's password.
  • Consent per action. A broad, standing permission may not be enough. Large, unusual or first-time transfers could ask the customer to confirm.
  • Capability limits. Caps on amounts, frequency and payees limit what a confused or compromised agent can do.
  • Speed. Agents could pause or ask a human before moving large balances, though a pause helps most if it is common practice.

A speech is not a rule. It does put a Fed governor on record treating unverified agent actions as a current problem, and agent speed as a stability question.

What to watch

  • Whether the Fed details the bank reports, such as how many incidents occurred and whether customers lost money.
  • Whether examiners apply the OCC's non-binding Nov. 4 supervisory framework ↗ for bank AI in practice. It covers agent identity, delegation, velocity limits and revocation.
  • Whether banks adopt a shared consent method such as Poppy ↗, Sierra's draft protocol, announced with Meta and others, which lists Bank of America and Wells Fargo among 35 new design partners.
  • Whether agent-driven outflows show up in bank liquidity planning and stress scenarios.

For now, the clearest signal is the present-tense report, not the run scenario: by Barr's account, banks are already seeing agents they cannot verify as genuine, approved by the customer, or following the customer's instructions.

Evidence note: Barr's remarks are quoted from the speech text published on the Federal Reserve's website. The bank reports he describes are his account; AINewsIQ has not seen them. Other reporting is linked inline and attributed to each outlet; American Banker's Oct. 27 story is cited from its published summary. [Bloomberg's report on the speech](https://www.americanbanker.com/external-article/fed-s-barr-says-ai-agents-could-fuel-future-bank-runs), republished by American Banker, said agents could speed "destabilizing stampedes out of banks."

This article was researched and drafted with AI assistance, then fact-checked and edited by Paul Woll, who is responsible for its content. Sources are linked inline.