Verifying AI agents’ identities may separate the rogues from the reliable.
Reported incidents of rogue artificial intelligence agents gaining unauthorized access to customer bank accounts have raised questions about their use in corporate payments.
Banking giants, JPMorgan Chase & Co., Bank of America Corp., Citigroup Inc., Wells Fargo & Co., and Morgan Stanley & Co. have either restricted access of third-party AI agents or blocked their access completely, citing the difficulty in knowing whether they have difficulty detecting which activity is authorized by the customer or a result of a rogue AI agent.
Without verifiable digital credentials to identify AI agents and their mandates, limiting or eliminating their access to bank systems is the right call, said Andrew Shikiar, CEO of the authentication standards organization FIDO Alliance.
“It’s going to be tough to address these issues that have already played out in ecommerce and fintech around trust, liability, and authorization,” he said. “With humans, it’s hard. With agents, it’s even harder.”
Digital Identities
Verifiable digital identities have been in operation for several years, but their implementation is fragmented. Sweden and Norway issue BankIDs, which provide digital credentials for banking and access to social services. Separately, the European Union has mandated that all member states issue citizens an EU Digital Identity Wallet by the year’s end, with acceptance by banks and other large platform providers by 2027. The EU Commission, however, has reported doubts that every member will meet the deadlines.
Meanwhile, in the Asia-Pacific region, Singapore offers its Singpass, which lets residents access online government and business services using biometric or two-factor authentication.
For the U.S., the federal government has left digital identity regulation to the states. Arizona, California, Colorado, Georgia, Maryland and Utah offer mobile digital licenses that can be stored on a person’s cellphone.
“Most people don’t know what to do with them,” said Shikiar. “It is not super well-coordinated.”
The FIDO Alliance has been working with the International Standards Organization, the OpenID Foundation and the World Wide Web Consortium to create a trusted and interoperable identity ecosystem that will let individuals store, manage, and share digital versions of their credentials like a driver’s license or passport.
“Once I have that, my agent can point to this signed, highly verifiable individual that authorized the agent to do X, Y, and Z with certain parameters and be audited, tracked, and verified,” said Shikiar.
Agent Control
Maintaining up-to-date controls around AI agents is of critical importance to the corporate treasury, and the company in general.
“Control testing examines individual steps, such as transaction approval, user access, or checks that a payment matches an authorization,” said Chris Dimitriadis, global strategy officer at ISACA, a professional association focused on IT governance, cybersecurity, risk management and auditing. “With AI agents, auditors must also test the integrity of the end-to-end sequence, the agent’s authority, and the transition between actions. The key difference is testing the full execution path, including the unexpected paths.”
Any time a company updates an AI model, deploys a new tool connection, changes an AI prompt, or gives an AI agent a broader mandate, it should reassess its control functions, he added. “I would establish a risk-based AI-agent change trigger rather than treating every update equally. Additionally, I would also establish an AI maturity and capability assessment framework like CMMI AIM to gain more confidence and reduce uncertainty.”
For now, Wall Street’s decision to halt agentic access to client accounts is a temporary, effective response until agent management catches up with the technology.
Rob Daly covers the economy and fintech. He can be reached at rdaly@gfmag.com