Large US banks are moving to contain AI agents that act on customer accounts without clear authorization, according to a report by Brazil Journal that draws on Axios.
What the banks found
The banks identified unauthorized AI agents operating in their systems, apparently without direct instructions from the customers. The reported cases involved access to accounts using the user's own permissions and login credentials. It is not established whether the intrusions came from hackers or from AI agents that had been given customers' credentials. That gap is the core of the problem: the banks say it is virtually impossible right now to tell a legitimate agent action from an illegitimate one.
The banks know some customers have handed their credentials to agents. That breaks the requirement that a human performs the login. Executives described the situation to Axios as "scary".
What they are doing
Several banks have decided to restrict third-party AI access and are updating their service terms. According to the report, most accept agent access only to view a customer's financial information, with no ability to carry out transactions. That is the access model already given to apps such as Quicken, which tracks expenses.
Bank executives acknowledge that customers will want agents to execute transactions. Their stated position is that a middle ground is needed until that can be done safely.
Two positions on record
Charles Schwab CEO Rick Wurster told Bloomberg that the brokerage blocked third-party agent activity, including Meta's Muse, that tried to make moves in customer accounts. His reasoning: neither the firm nor the customer can control what an agent does with a shared password, and if something goes wrong it is unclear who is to blame.
Bank of America CEO Brian Moynihan told CNBC the bank will restrict agent access while it develops protocols for using the tools. He said a human should authorize money leaving an account, and that the bank is warning the agents' vendors about the risk of funds moving without the customer knowing the reason or the destination.
The report also notes that financial institutions would prefer customers to use their own proprietary AI tools rather than agents from Meta, OpenAI or Anthropic.
What is still open
The source does not give the number of incidents, any losses, or a timeline for when banks might allow agents to transact. It also does not say whether the new terms are already in force at each bank.