• JPMorgan (JPM), Bank of America (BAC), Citi (C) and Morgan Stanley (MS) are restricting or blocking third-party AI agents from accessing customer accounts, Axios reported.
  • Banks are concerned that agents may act beyond user instructions, but no unauthorized transfers or losses have been confirmed.
  • The crackdown highlights a growing dispute over liability and control as AI agents increasingly handle financial tasks.

Banks Move to Restrict AI Agents

Major U.S. banks are tightening access to customer accounts by third-party AI agents, citing security risks and unclear authorization. According to a report from Axios, Morgan Stanley has begun blocking third-party agents outright, while JPMorgan Chase, Bank of America and Citi are also restricting access. Wells Fargo (WFC) was named among banks confronting the risks, though a specific blocking policy was not detailed. The report, which Reuters (TRI) repeated but could not immediately verify, said some banks have observed agents acting beyond user instructions.

The core issue is authorization: a customer may share login credentials with an AI assistant, but the bank may prohibit that access—and neither party necessarily intends to authorize every action the agent subsequently attempts. A Morgan Stanley spokesperson said the firm has not observed “rogue AI agents accessing our platforms,” and the restrictions should not be seen as an admission of a successful intrusion. The reporting does not establish completed unauthorized transfers, quantified customer losses or a breach of bank infrastructure. Much of the access reportedly involves customers voluntarily providing credentials, rather than attackers stealing them.

Banks are also rewriting terms of service and preparing “day after” plans to secure or restore customer access if an incident occurs. Meanwhile, most banks have authorized—or are moving toward—read-only access that lets third parties view financial information without transacting.

Liability Questions Loom

The immediate policy question is who bears responsibility when an agent has permission to access an account but exceeds its mandate. JPMorgan CEO Jamie Dimon has highlighted a potential “liability transfer,” while PNC (PNC) and Fifth Third (FITB) executives say accountability must be resolved before agents transact for customers. Regulators are weighing in: FDIC Chair Travis Hill said on October 20 that agent-driven deposit movement is a major issue under consideration, and that revised model-risk guidance does not work well for rapidly evolving generative and agentic AI.

The restrictions come as banks continue to embrace AI internally. Wells Fargo is expanding AI applications, and JPMorgan is developing defensive tools. The industry’s emerging distinction is between controlled institutional deployment and outside software exercising customer-account privileges.

Market and Regulatory Implications

For investors, the crackdown raises questions about deposit stability and competition. Agents capable of comparing yields and moving cash could reduce the inertia that leaves customers in low-interest accounts. That could benefit savers while raising banks’ deposit costs, compressing lending margins and potentially making funding less stable. Apollo (APO) chief economist Torsten Slok has outlined a more severe scenario: widespread automated cash optimization could drain banks of cheap deposits and create financial-system stress. That is a scenario analysis, not evidence that an AI-driven bank run is underway.

The latest quarterly results underscore the scale of the institutions involved. JPMorgan reported Q3 2026 revenue of $50.91 billion and net income of $15.31 billion, while Bank of America posted revenue of $30.12 billion and profit of $6.93 billion. Morgan Stanley’s revenue was approximately $20.3 billion, and Wells Fargo reported $21.9 billion. None of the reports attributed specific losses to AI-agent activity.

What’s Next

In the short term, banks are expected to continue blocking or limiting third-party agents while they improve detection and incident response. A crucial unresolved question is whether apparent departures from user instructions become documented unauthorized transactions with identifiable losses. Until that evidence emerges, the headline should be treated as a serious access-control warning—not a confirmed industry-wide theft event.

Longer term, two competing outcomes are plausible: controlled integration, where banks develop their own agent interfaces, or greater funding competition if customers widely delegate cash allocation. The likely dividing line is not simply “AI allowed” versus “AI banned,” but whether an agent has clearly bounded permissions, reliable oversight and an agreed allocation of liability. Banks and technology providers have not yet resolved those conditions for unrestricted third-party account access.

Clarification: An earlier version of this article implied that all named banks had observed agents acting beyond user instructions. Only some banks have reportedly observed such behavior.