U.S. Banks Restrict AI Agents as Survey Shows 51% of Banks Testing Them
Key Facts
Banks have detected activity by AI agents that appears independent of users' instructions, Axios reported on October 22, citing people familiar with the matter, amid banking restrictions on the tools' access to customer accounts. The Financial Times reported that Citi had joined banks restricting agents' access to accounts. A Morgan Stanley spokesperson, however, said the bank had not seen instances of "rogue AI agents accessing our platforms." In JPMorgan's official earnings transcript, the chief financial officer described sharing login credentials and screen scraping as unsafe methods. "It's unsafe," he said of the approach. He added that the bank blocks, as far as it can, services that rely on those two methods to access accounts.
JPMorgan's warnings put customers' liability for agents' actions at the heart of the risks associated with their use. Jamie Dimon warned in the bank's official earnings transcript that liability shifts to customers when they use these tools, saying: "There is a shift in liability." In payments, Zach Anderson, the bank's head of data and analytics for payments and global banking, described getting agents to execute payments reliably and securely as the difficult challenge. "It's extremely difficult to do that reliably and securely," he told Axios. A note published by the International Monetary Fund in April 2026 examined risks involving the traceability of authority delegated to agents, the opacity of their behavior and the interconnectedness of their actions. The note also listed cybersecurity, legal questions and unresolved liability issues among the risks associated with their use in payments.
The restrictions on external access to accounts coincide with banks testing agents within their own operations. Reuters reported in July 2026 that a KPMG survey conducted in June found 51% of banks were testing AI agents. "We use agents internally," Dimon said in JPMorgan's earnings transcript. As an example of oversight of such use, Reuters quoted BNY's chief executive as saying that a digital employee within the bank reports to a human manager responsible for its training. For payments on users' behalf, Mastercard's "Agent Pay" program requires trusted agents to be registered and verified before they are allowed to pay. The company said in its April 2025 launch statement that, once registered and verified, agents would be able to make secure payments on their users' behalf.
Beyond account security, JPMorgan linked competition from agents to money flows seeking yield. The bank also pointed to competition from fintech companies over the customer experience. Discussing that competition, it said one aspect involved "yield-seeking flows." The chief financial officer said the "Smart Cash" service had begun operating as part of a limited pilot aimed at a particular customer segment, describing it as a "targeted limited pilot." At Citi, Jane Fraser said in the official earnings transcript that she was not concerned about the risk of deposit outflows in the Services business. "I'm not concerned about the risk of deposit outflows in the Services business," she added.
On oversight, Reuters reported in October 2026, citing sources, that U.S. regulators were asking banks about safeguards to prevent agents from exceeding their authority. Fraser said Citi uses advanced models within a controlled framework to identify and address vulnerabilities quickly and at scale. She added that the bank works with peers to coordinate defenses. Dimon said JPMorgan was building tools to protect customers from agent-related risks. In Europe, the deadline the European Central Bank gave banks in July is approaching, with banks required to prepare plans to counter AI-powered cyber threats by the end of October, according to Reuters.