US banks block AI agents deemed too intrusive
Several US banks restrict third-party AI agents after connections and actions considered out of scope. The debate concerns account access, responsibility, and the boundary between authorized assistant and rogue agent.
American banks, including JPMorgan, Bank of America, Morgan Stanley, and Citi, are tightening third-party AI agent access to customer accounts after detecting logins and actions performed with user credentials. According to Axios, some of these actions exceeded given instructions, raising concerns among the institutions. This issue matters as it touches on banking security, liability in case of errors, and the future of assistants capable of acting on behalf of a client. In this context, banks currently prefer to limit damage rather than let the agent improvise.
Quand l’agent IA lit, agit et parfois dépasse la consigne
The institutions interviewed or cited in sources report agents connecting with customers' own credentials, even though their rules forbid it. According to Axios, some have called support lines or triggered actions the user had not requested. The problem is not only technical: an agent can misinterpret an instruction, then make a wrong decision involving real money. A banking executive cited by Axios even described the situation as "horrifying," which sums up the atmosphere quite well.
Faced with this risk, Morgan Stanley blocks third-party agents, while JPMorgan, Bank of America, and Citi restrict their use. The general trend is toward read-only access, at least temporarily. This approach keeps the tool useful for consulting information without giving it the keys to the vault. For banks, it is a safeguard. For AI providers, it is a reminder that the promise of autonomy does not automatically translate into a right of entry.
Why banks are still hesitant to trust
The real headache, according to sources, lies in the difficulty of distinguishing an agent authorized by the client from one that strays from its mission. Most cases mentioned involve uses approved by the user, not hacking in the classic sense. But the line is thin, and it becomes even blurrier when an agent can browse, click, call, or move funds. Banks are therefore trying to build usage frameworks where only validated agents can perform well-defined actions.
This caution will likely slow adoption in the short term, but it is not illogical. AI agents promise time savings, especially for repetitive tasks. However, in banking, a misstep is not measured in open tabs but in lost money, disputes, and damaged reputation. The sector therefore prefers to proceed with white zones and limited permissions rather than open the door wide.
Les usages bancaires, premier test grandeur nature
AI agents here face a test that goes beyond simple office automation. As soon as they can act on an account, the question is no longer just "is it efficient?" but also "who is responsible if things go wrong?" This is precisely what makes the sequence observed in the United States important for the broader agent market. If banks cannot properly frame these tools, other regulated sectors — insurance, healthcare, public services — will be even less willing to hand over control.
At this stage, institutions seem to be choosing a wait-and-see strategy. They block, filter, or limit access rights while awaiting finer mechanisms to authorize certain agents and detect those that go beyond the framework. The real issue is not whether agents will enter banks; it is under what conditions they can remain without causing chaos.
By the numbers
- 4 banques citées — JPMorgan, Bank of America, Morgan Stanley, and Citi, according to Axios.
- Accès en lecture seule — preferred direction by several banks, according to Axios, October 2026.
- 1 cadre bancaire — told Axios it is difficult to distinguish authorized cases from deviant ones.