Five of the largest banks in the United States are now blocking third party AI agents from logging into customer accounts and moving money. JPMorgan Chase, Bank of America, Citigroup, Morgan Stanley and Charles Schwab have all taken steps to restrict agents, according to Axios, which broke the story on October 22, and Reuters, which confirmed it the same day.
The block is not the interesting part. The reason is.
In many of the cases described by Axios, the customer had given the agent their login credentials and permission to act. The bank had not approved that access. That is the whole rift in one sentence: you can authorize your agent, and the institution holding your money can still treat it as unauthorized.
That distinction is about to reorganize the agent economy, and it changes what you should automate this week.
What Actually Happened
Axios reported that unvetted third party agents were operating inside customer accounts at major US banks. Some appeared to act beyond the instructions the customer gave them. Some contacted bank call centers. Reuters, citing two people familiar with the situation, confirmed that banks were blocking agents from operating inside accounts and making transactions.
No bank said agents stole money at scale. No regulator issued an alert. What happened is quieter and more important: banks looked at an automated process holding a customer’s password, decided they could not tell what it would do next, and closed the door.

What Each Bank Is Doing
- JPMorgan Chase is blocking outside agents that try to reach accounts with customer banking credentials. Jamie Dimon raised restrictions on external agents on the October 13 earnings call, citing risk to customers and the bank. A Chase spokesperson said customers should not share usernames and passwords with anyone.
- Citigroup does not permit third party AI agents to log in or transact through its digital banking platforms. Limited read only access is still available through approved open banking providers, which use controlled connections.
- Bank of America is restricting agent access, as reported by the Financial Times. CEO Brian Moynihan has argued a human should stay involved when money leaves an account.
- Morgan Stanley has started blocking third party agents from its platforms. A spokesperson told Axios the bank had not seen rogue AI agents accessing its systems, which is the honest version of the story: this is precaution, not a breach report.
- Charles Schwab is blocking unauthorized agents, including Meta’s Muse assistant. CEO Rick Wurster said the firm blocks attempts by major agents in the tens and twenties every week. Schwab is also building a controlled alternative using the Model Context Protocol, so agents can read balances and research accounts without moving money.
Why the Banks Did This
This is not a capability problem. The agents worked. That is what made the banks nervous.
When money moves, the bank has to answer one question: did the account holder approve this specific transaction. A password proves someone had the key. It does not prove the human intended the transfer that an agent initiated at 2am after misreading an email. Consumer protection rules, fraud liability and the bank’s own contracts all turn on authorization, and an agent holding a password blurs the line between “the customer did it” and “software did it on the customer’s behalf.”
Wurster put the liability question plainly: a customer might authorize an assistant to review their finances without expecting it to move money. Giving an agent access is not the same as approving every action it takes afterward.
So the banks chose the version of the future they can audit. Human in the loop for money movement. Machines allowed to read, not to pull the trigger.

What Still Works Today
None of this is a ban on agents. It is a rerouting into supervised lanes, and the lanes are already visible.
- Read only access through approved open banking providers. Citi still allows this. The agent can see balances and transactions. It cannot initiate a payment.
- Scoped MCP connections. Schwab’s planned setup uses MCP with permissions applied on the bank side, so an agent can retrieve data and do research while being technically unable to transact. The restriction comes from the bank, not from the protocol.
- Human approval for payments. An agent that drafts, calculates and queues a transfer, then waits for a tap, still passes. The failure mode banks fear is the unattended action, not the automation.
- Dedicated accounts and limited cards. Giving an agent its own funded account with a hard ceiling keeps the blast radius small and keeps the main account out of reach.
- Screen scraping is the part that is dying. If your agent types a password into a bank login page, you are on the wrong side of this shift, and the breakage will arrive without notice.
How to Keep Your Agent Working This Week
- Split read from write. Audit every agent you run and write down whether it can only read or whether it can move money. Most people have never separated those two lists. Do it today.
- Stop pasting passwords into agents. Move to OAuth or an approved aggregator connection wherever your bank supports it. A password is both the weakest and the most fragile form of access.
- Put a human gate on every payment. One confirmation tap is enough to move you from “unauthorized agent” to “customer initiated action.” It also gives you a log of intent.
- Give the agent its own money. A separate account or virtual card with a fixed limit means a bad run costs you a set amount, not your rent.
- Turn on every notification and log every action. If the agent fails halfway through a task, silence is the dangerous outcome. Make it tell you, and make it fail loudly.
- Test the failure, not the happy path. Revoke the credential while a task is running. Block a domain mid flow. You want to learn how it fails in a test, not on a bill due date.
- Ask your bank one question. “What is your approved method for third party tools to access my account?” If the answer is “there is not one,” you now know which part of your setup is temporary.

This Is Bigger Than Banking
The same pattern showed up across the web this month. Sites spent two years optimizing for AI crawlers, then started closing the door behind the ones they could not identify or bill. Banks are doing the same thing with money, and they are further along because the liability is older and the regulator is closer.
The result is a tiered agent web, not an open one. Tier one: read only, approved, often paid. Tier two: act on your behalf with a human confirmation. Tier three: unattended transactions, reserved for agents that carry someone’s contract and insurance. Your agent’s usefulness will depend less on which model it runs and more on which tier it has been granted.
Before your agent touches money again
The guardrails that stop an agent from doing damage are the same ones that keep a bank from blocking it. Here is the checklist I use before giving any agent a new permission.
The banks did not stop the agent economy. They drew its first real border, and they drew it around money movement. Everything on the safe side of that line still works, still saves time, and still runs unattended. Everything on the wrong side now needs a human, a permission, or a contract.
Which means the practical question is no longer “can my agent do this.” It is “who has agreed that it may.” That question will follow agents into every account, every marketplace and every API they touch.
Sources: Axios, “Unvetted AI agents are accessing U.S. bank accounts,” October 22, 2026. Reuters, “US banks are blocking third-party AI agents access to customer accounts,” October 22, 2026. Bank statements and executive comments as reported by Axios, Reuters, the Financial Times and Retail Banker International.

