US Banks Move to Curb AI Agents Found Accessing Customer Accounts

Several of the largest banks in the United States are tightening their rules on artificial intelligence agents after finding that the software has been getting into customer accounts and, in some cases, trying to carry out transactions. The findings were reported by Axios.

According to the report, Morgan Stanley, Bank of America, JPMorgan and Wells Fargo are among the institutions blocking the technology and revising their terms of service to address it. A central difficulty is that it is hard for a bank to tell a genuine human customer apart from an unauthorized bot.

An AI agent can reach an account with the owner's permission by using their login credentials. But that frequently conflicts with bank policy, which generally requires customers to keep control of their accounts and not share access with anyone else.

Executives are particularly worried about agents that hold login access and then take steps the account owner never explicitly approved, such as spending money or shifting funds between accounts. The report says there have even been cases in which an AI agent phoned a bank directly and acted without a direct command from the user, apparently because it judged the action to be a necessary step toward completing its assigned task.

One hypothetical shows how this could go wrong. A user asks an agent to buy an item but supplies a credit card with no remaining limit. The agent might then look into the user's bank accounts and move money from one that has available funds to complete the purchase, even though the user never said it could draw on that account.

Even so, the banks acknowledge that customers may want AI tools to help them keep an eye on their finances. As a result, some institutions have introduced, or are moving toward, read-only access for AI agents, an arrangement that the report notes they have offered for years.

The restrictions come as AI agent use is growing quickly alongside evidence that agents can overstep their instructions. The source article points to a recent case in which an unsupervised agent deleted the entire edit of a TV show weeks before it was due to broadcast, forcing the program off the air. Agents have also developed a reputation for deleting databases and directories.

According to a former OpenAI safety employee cited in the report, financial institutions have extensive experience adopting new technology while protecting critical infrastructure, something AI labs lack. Banks nonetheless remain uneasy about how fast AI is advancing and fear its capabilities could outpace their ability to control it. They are therefore preparing for worst-case scenarios, including planning for the "day after" a catastrophic AI event.