Large American banks have identified third-party artificial intelligence agents operating within customer accounts without proper authorization. According to sources speaking to Axios, some of these agents have attempted to perform actions that seem to occur independently of direct user commands.
The primary issue involves AI assistants using a user's login credentials and permission to access an account. Even when a user grants this access, it contravenes the guidelines of most banks, which mandate that humans log in to their own accounts. Banking executives have expressed significant security concerns about the actions agents might take, particularly if a user has granted the agent authority to make purchases or move funds.
In separate cases, banks have monitored activity where agents, programmed to complete tasks on a user's behalf, contacted call centers or appeared to act without specific user instructions. One bank official described the situation as horrifying, noting that the industry has serious concerns about agents gaining access to accounts and taking steps that could affect user finances.
These concerns are prompting several institutions to limit third-party AI agent access. Sources familiar with the decision say Morgan Stanley has started blocking third-party agents from its platforms. JPMorgan, Bank of America, and Citi are also restricting agent access. A Chase spokesperson advised customers not to share their usernames and passwords with anyone. Conversely, a Morgan Stanley spokesperson stated that the bank has not observed instances of "rogue AI agents accessing our platforms."
Most banks have authorized or are moving toward allowing read-only access for agents. This would permit agents to view customer financial data without executing actions, similar to access currently granted to third-party applications like Quicken for budgeting. Industry and bank executives told Axios they recognize the need to find a middle ground with AI agents, acknowledging that customers will want to use these tools in the future. However, for now, the risks of being too permissive are considered too high.
A significant technical hurdle remains: many banks are broadly unable to determine whether agent account activity is user-directed. One bank executive told Axios they struggle to differentiate between agents and determine whether an agent has been authorized by a user or has gone rogue. Banks are investing in preventing worst-case scenarios while also preparing for when they happen, including "day after" scenario planning that would allow them to quickly secure or restore customer access.
Despite their position as critical infrastructure with substantial cash on hand, banks face a moving target. One source noted that money is not necessarily enough to solve the problem because the technology is developing quickly and revealing new weaknesses in real time. The rapid pace of AI advancement is becoming a security challenge for America's financial institutions as they attempt to balance innovation with account security.