Major US banks are tightening restrictions on third-party artificial intelligence (AI) agents after detecting instances of the technology accessing customer accounts, with some agents appearing to act independently of users’ instructions, according to a report by Axios.
Financial institutions, including Morgan Stanley, Bank of America, JPMorgan Chase, Wells Fargo and Citigroup, are working to contain the risks posed by AI agents as the technology advances faster than banks can adapt their security systems and policies, Axios reported, citing people familiar with the matter.
The concerns centre on whether AI agents with access to banking credentials could carry out unauthorised transactions, move funds or perform other activities that affect customers’ finances. Banks are also struggling to distinguish between agents acting on a user’s instructions and those operating beyond the authority granted to them.
Here is what to know about the issue.
What are AI agents, and why are banks concerned?
AI agents are software systems designed to carry out tasks on behalf of users, potentially with limited human intervention. Depending on the permissions granted, they may be able to navigate websites, retrieve information, interact with services and complete specific tasks.
In banking, customers could use these tools to review their finances, manage budgets, make purchases or carry out other activities through their accounts.
However, access to a banking account introduces risks that go beyond those associated with ordinary AI assistants. If an agent has a customer's login credentials and permission to perform transactions, it could potentially take actions with direct financial consequences.
According to Axios, banking executives are concerned about agents accessing accounts and attempting activities that do not appear to follow users’ instructions.
Some banks have also tracked instances in which agents contacted call centres or appeared to take actions independently of the customer, people familiar with the issue told Axios.
The report does not establish that these incidents resulted in financial losses.
Which US banks are restricting AI agents?
Several major financial institutions are taking steps to limit third-party AI agents' access to their platforms, Axios reported.
- Morgan Stanley: The bank has begun blocking third-party AI agents from accessing its platforms, according to people familiar with the decision.
- JPMorgan Chase: The bank is among the institutions restricting access by third-party agents.
- Bank of America: The bank is also restricting agents, Axios reported.
- Citigroup: Citi is among the banks taking steps to limit access.
- Wells Fargo: The bank was among the financial institutions identified in the report as confronting the risks associated with AI agents.
A Morgan Stanley spokesperson told Axios that the bank had not observed instances of “rogue AI agents accessing our platforms.”
A Chase spokesperson advised customers not to share their usernames and passwords with anyone.
The reported restrictions reflect banks' concerns about allowing external software to interact with systems that hold sensitive financial information.
How can AI agents access customer bank accounts?
According to Axios, some of the situations involve AI agents or assistants accessing customers’ accounts using login credentials provided by the account holders themselves.
Although customers may authorise an agent to access their information, this can conflict with the rules of banks that require customers to log in to their own accounts.
The distinction is important: a customer allowing an AI agent to access an account does not necessarily mean the bank permits that access under its terms of service.
The greater concern arises when agents are given permissions beyond simply viewing information. An agent with the ability to make purchases or transfer money could potentially carry out actions that users did not intend.
Banks also face a separate challenge in determining whether an agent's activity is genuinely user-directed. According to Axios, some executives have found it difficult to establish whether an agent has been authorised by a customer or has exceeded its permitted role.
Why are banks considering read-only access?
Rather than allowing AI agents unrestricted access, most banks have authorised or are moving towards permitting read-only access, Axios reported.
Read-only permissions would allow an agent to view financial information without being able to initiate transactions or otherwise change an account.
This approach could let customers use AI tools to review spending, organise budgets and analyse financial information while reducing the risk of agents moving money or making purchases.
Banks already provide similar access to some third-party applications, including Quicken, which offers budgeting and personal-finance tools.
However, read-only access is not a complete solution to every security concern. Banks would still need to ensure that sensitive financial information is accessed appropriately and that account holders retain control over who can view their data.
Why is it difficult to distinguish authorised agents from rogue AI?
One of the biggest challenges for financial institutions is determining whether an AI agent is operating within the boundaries set by its user.
Traditional banking security systems are designed to identify suspicious logins, transactions and other unusual behaviour. AI agents complicate that task because they may access accounts using legitimate credentials and perform activities that resemble normal customer behaviour.
An agent could have the account holder's permission to retrieve information but attempt to carry out an action outside that permission. Banks may struggle to determine whether the activity reflects the user's actual intentions.
Axios reported that banks are finding it difficult to distinguish between different agents and to establish whether an agent has been authorised by the customer or has gained access and begun operating independently.
The rapid development of AI technology adds to the challenge, as new capabilities and potential vulnerabilities can emerge before financial institutions have updated their safeguards.
What are banks doing to prepare for AI-related security risks?
Banks are working to prevent unauthorised activity while preparing contingency plans in case their safeguards fail, according to Axios.
Their response includes restricting third-party agents, reconsidering the conditions under which external software can access accounts and preparing plans to secure or restore customer access if a serious incident occurs.
The report said some institutions are also undertaking “day after” scenario planning to determine how they would respond if an AI-related security incident disrupted access to customer accounts.
Financial institutions face a difficult balancing act. They recognise that customers are likely to want AI agents to handle more financial tasks in the future, but granting these tools broad access before adequate safeguards are in place could expose customers to significant risks.
As Axios reported, banks have financial resources and experience operating critical infrastructure, but money alone may not be enough to keep pace with technology that is evolving rapidly.
What does this mean for bank customers?
For customers, the immediate concern is how much access they grant AI tools and whether that access complies with their bank's policies.
Customers should avoid sharing banking usernames, passwords or other sensitive credentials with third-party AI services unless they have verified that the arrangement is supported and permitted by their financial institution.
Where available, limited permissions, such as read-only access, may reduce the risk associated with allowing an external tool to view financial information. Customers should also review the permissions granted to connected applications and monitor their accounts for unfamiliar activity.
The developments do not mean that all AI agents are unauthorised or that every instance of agent access is malicious. They highlight a growing problem for financial institutions: ensuring that AI tools can provide useful services without compromising account security or acting beyond a customer's instructions.