The alarm began as major U.S. banks detected third-party AI agents entering customer accounts, usually after customers handed over their own login credentials. Executives told Axios the core problem was not simply bot traffic: banks often could not determine whether an agent was acting on a customer’s instructions, had exceeded them, or had obtained access improperly.¹

That uncertainty quickly produced a defensive response. Morgan Stanley began blocking third-party agents, while JPMorgan, Bank of America and Citi also moved to restrict access. Banks fear an agent with authority to make purchases or transfer money could take steps a customer never intended. Morgan Stanley said it had not seen “rogue AI agents accessing our platforms,” but the industry’s inability to distinguish legitimate automation from something more dangerous has been enough to force a clampdown.¹

The banks’ immediate advice is blunt: do not share passwords. JPMorgan said, “Under no circumstance should customers share passwords with any person or thing. Period,” pointing instead to secure open-banking services that can provide limited access without surrendering account credentials.²

Yet the restrictions have opened a second argument over what, exactly, is being blocked. David Sacks, the former White House AI adviser, said there is “a world of difference” between customer-authorized agents and truly rogue systems that could amount to cyberattacks. He argued that much of the activity described appeared to fall into the former category, warning against treating every user-directed agent as a security incident.David Sacks (DavidSacks)³

Banks are not rejecting automation outright. Many already allow—or are moving toward—read-only connections, similar to tools used for budgeting. But with AI agents evolving faster than bank controls, the industry is preparing not only to prevent failures but also for the “day after” a major cyber disruption, including plans to restore access and reassure customers.²