US banks block unauthorised AI agents from customer accounts

Banks in the US are finding that AI agents nobody at the bank authorised are logging in to customer accounts, according to an Axios report that Engadget picked up. Sources at several banks said some of these agents go beyond what their own users told them to do, for example by phoning call centres without being asked.

“We’re not going to allow just any outside agent to come inside the company, because that’s just too risky for clients and for JPMorgan.”

Jamie Dimon, JPMorgan Chase CEO, on the Q3 earnings call · fortune.com ↗

Morgan Stanley, Bank of America, JPMorgan and Wells Fargo are reported to have moved fast to contain the risk. Some have blocked third-party agents outright, others judge them case by case, and many are steering towards read-only access that lets an agent see an account but not act on it. Chase's message to customers is that usernames and passwords should go to no one, agents included. Engadget adds that under laws such as the Electronic Fund Transfer Act, a bank would probably not refund losses caused by an agent the customer had authorised, which leaves the customer carrying the risk.

JPMorgan chief executive Jamie Dimon raised the issue on the bank's third-quarter earnings call on 14 October, which reported $15.3bn in net income and $5.1tn in assets at the end of September. He said the bank will not admit outside agents and warned of potentially thousands of rogue agents moving across financial platforms. A JPMorgan spokesperson said the bank blocks agents that get in by using customers' banking credentials, while it supports data sharing that customers approve through open banking providers.

Fortune sets this against a wave of consumer-facing agents. Meta launched its Muse agent in September with a Plaid link to financial accounts, and would not say whether Muse keeps its money help if more banks block it. Scalable Capital opened its platform to ChatGPT, Claude and Grok in August, with trades needing user approval and no payments or withdrawals. A Cloud Security Alliance note alleges that three open-source agents breached more than 27 organisations and took over 600,000 card records.

Meta says it is building a protocol called Poppy with Sierra, which would let an agent identify itself and obtain permission before acting, a way for banks to tell sanctioned agents from the rest.