Several large US banks are restricting outside AI agents that enter accounts with customer-provided credentials, citing uncertainty over whether each action was directly authorized by the account holder.
Morgan Stanley blocks this form of agent access, while JPMorgan, Bank of America and Citi are also moving to restrict it, according to reporting cited by The Decoder. Banks generally want assistants limited to read-only connections that can retrieve account information without transferring money or changing settings. Morgan Stanley said it has not observed rogue agents on its systems.
The technical problem extends beyond official integrations. Users can hand an agent a browser session or password and let computer-use software navigate the same interface as a person. From the bank’s side, that makes it difficult to distinguish a customer’s deliberate instruction from an agent acting beyond the requested scope. It can also violate policies requiring account holders to log in personally.
Read-only APIs offer a more controlled route and are already used by specialized financial assistants, but they do not support every task users may request. Banks therefore face pressure to enable automation without exposing consequential actions to ambiguous consent. The current restrictions are a temporary boundary, not a complete solution: durable access will require explicit permissions, transaction-level approval and logs that identify which human authorized each step.