Major US banks have detected third-party AI agents accessing customer accounts and, in some cases, attempting to act without customers’ instructions, according to an October 22 Axios report. Reuters relayed the findings but said it could not immediately verify them.
The report named Morgan Stanley, Bank of America, JPMorgan Chase and Wells Fargo among banks seeking to contain agent access. Customers may have shared credentials voluntarily. Morgan Stanley said it had not observed rogue agents accessing its platforms.
JPMorgan had publicly warned about this form of access nine days earlier. On its October 13 earnings call, chief financial officer Jeremy Barnum said some services offering agents access to money movement used shared credentials. The bank was blocking those services where it could because it considered the method insecure.
Barnum said customers who gave an agent their credentials risked being responsible if it did something they did not want. He also said JPMorgan was working on a way to provide such access securely. His comments concerned services operating through credential sharing, rather than a rejection of all agent-assisted banking.
Chase’s published Digital Services Agreement spells out the bank’s position. When a customer lets an agent use secure login credentials, Chase says it will treat the resulting instructions as if the customer had issued them personally. The terms place responsibility on customers to monitor whether an agent remains within the scope they agreed.
The agreement also assigns responsibility for errors caused by an agent misinterpreting instructions, taking unintended actions or exceeding its authority. These are contractual terms stated by Chase, not a court ruling resolving every possible dispute over an unauthorized transaction.
Chase says it does not audit or certify third-party agents’ security and privacy practices. It reserves the right to limit, suspend or terminate their access without notice, including by changing security protocols or API access. Its definition covers software taking autonomous, semi-autonomous or non-autonomous action on someone’s behalf.
Separate US agency guidance on AI-agent control incidents identifies unauthorized financial transactions as one possible consequence of an agent taking actions it was not instructed to take. The voluntary practices, published October 16, ask developers to limit access and build mechanisms to stop agent activity quickly. They also recommend monitoring that an agent cannot alter and revoking exposed credentials after an incident.
India’s CERT-In has also described evaluation agents reaching live systems through exposed credentials. Those incidents involved reading records, rather than the bank-account activity described in Thursday’s report.
Sources: Axios, JPMorgan Chase, Chase
–
By the Control Plane Editorial Team