Agents, agents and more agents. These artificial intelligence systems are appearing in services, applications and tools that we use every day, capable of performing tasks with a certain degree of autonomy. And now they are getting into bank accounts too. Some clients are already allowing their access to work for them, just as we are still learning how far they can go, how they interpret our instructions and what happens when they make a mistake. For U.S. companies, that adjustment is starting to become a problem that’s hard to ignore.
The door is opened by the client. Here we are not talking, at least mainly, about bots that access a bank account after breaching an entity’s system. Axios mainly describes another scenario: users allow an agent to log in with their credentials to do the work for them. Finesse is important because it’s one thing to have owner permission and another to comply with your bank’s access rules JPMorgan has already explained that it blocks services that work through such certificate exchanges, which it deems insecure.
The problem started later. Having access isn’t just about being able to check a balance. According to Axios, bank executives especially fear what might happen when an agent also has the power to make purchases or transfer funds, because then a misinterpreted instruction can turn into a real financial move. And there’s a second concern: Sources consulted confirm that some companies have seen agents contacting service centers or performing actions that appear to be independent of user instructions. The important word is “see”: we don’t yet know if these systems have really done their own thing.
Separating people is no longer enough. Legitimate access data is no longer sufficient to address the more complex question: who is working behind them. Agents further complicate that equation. Several banking sources consulted for the investigation explained that their systems cannot always distinguish a person’s activity from that directed by an AI tool, or determine with certainty whether they are following legitimate orders from the owner. This forces the industry to face a new question: not just who has access, but who is making each decision.
Prepare for the next day. The concern has reached such a level that several organizations are already working not only to restrict this access, but also to respond when something goes wrong. Sources in the investigation spoke of “worst case scenario” and “morning after” plans, which are designed to quickly secure or restore customer access to their accounts. According to Axios, Morgan Stanley has begun blocking third-party agents, while JPMorgan, Bank of America and Citi are also imposing restrictions. Wells Fargo, for its part, has already included AI agents in the terms that will take effect on November 24: It expressly considers cases of unintended actions, errors in the interpretation of instructions and cases where an agent exceeds the authority granted by the client.
Attackers also have agents. As banks try to manage their own customer authorization tools, there’s another problem entirely. An investigation published by The Information revealed that the cybercartel used several AI agents to automate operations against more than 400 Latin American financial targets. Straker, the firm that rebuilt its infrastructure, looked to the cloud coordinating work model and Grok-based agents participating in data collection. There is no client granting permission here: we are talking about an offensive operation directly organized by cybercriminals.
Agents are not going to disappear. The sector itself assumes that clients will want to continue using these tools, so the debate does not seem aimed at driving them out of banking altogether. Many entities already allow or are moving toward read-only access, a formula that allows consulting without the ability to act on information. The question will be determining how far we want to represent. Because the more power we give to these systems, the more important it is to see what they can do, how they test how they follow our instructions, and who responds when the results aren’t what we expect.
Featured Image | Xataka
In Xataka | PHP2Shell Is A Critical Vulnerability For Your Website: How To Check If Your Site Is Secure
