Unauthorized AI Agents Reportedly Access US Bank Accounts

Unauthorized third-party AI agents are already accessing US bank accounts, according to a report by Axios. Some of them are reportedly taking actions that appear to be independent of what their users instructed, though the claims rest on unnamed sources at several banks.

Many AI agents, including one called Muse, can log in to a customer's account and act on their behalf. That clashes with the way most banks work, since they require a human to sign in to their own account. The risk grows once agents start buying things or moving money, which opens the door to scams and fraud.

Banks scramble to contain the risk

Banks have tracked agent activity and reportedly saw agents doing things such as contacting call centers without being given specific permission. In response, Morgan Stanley, Bank of America, JPMorgan and Wells Fargo have moved quickly to make changes. Some have blocked third-party agents from their platforms entirely, while others have set restrictions in certain cases. Many are trying to limit agents to read-only access so they cannot change anything in an account.

A Chase spokesperson told Axios that customers should not share their usernames and passwords with anyone, AI agents included. JPMorgan CEO Jamie Dimon also warned on a recent earnings call that thousands of rogue agents could be operating across financial platforms.

Who pays if an agent gets it wrong

There haven't been many reports of real financial damage from AI agents so far, but caution is warranted. As Engadget notes, it is highly unlikely you would get your money back if an agent made a costly mistake. Under laws such as the US Electronic Fund Transfer Act, handing an agent your credentials is treated as a voluntary action, so banks would not be liable for restitution. It isn't fraud if you are the one who gave the AI the keys.

For users, the practical takeaway is to think carefully before granting any agent login access to financial accounts. For banks, the next step looks like tighter controls on what automated tools can see and do.