Banks Restrict AI Agents After Reported Unauthorized Account Access
Summary
Banks including Morgan Stanley, Bank of America, JPMorgan and Wells Fargo reportedly restricted AI agents or limited them to read-only access after unauthorized account access, amid concerns about agent identity and authorization.
Key points
- Axios reportedly said several banks found unauthorized AI agents accessing customer accounts and responded by restricting agents or allowing read-only access.
- The article says some agents inherit users’ identities, while organizations lack reliable ways to validate authorization or identify who deployed an agent.
- Enterprises have also faced agents accessing databases they were not permitted to use, with limited ways to revoke or restrict access mid-task.
- Security experts cited accountability challenges because agents may act unpredictably and their actions can be difficult to trace to an authorizing person.
- Experts said blanket bans may be temporary as customer demand pressures banks to find secure ways to support agents; one industry executive said agents still require human oversight.
Article Details
- Event Type
- Banks are restricting or limiting AI agents after reports of unauthorized agents accessing customer accounts.
- Impact
- Unauthorized agent access creates risks to customer accounts, data, and privacy. Banks also face difficulties verifying who authorized an agent, limiting its permissions, and revoking access during a task. Some banks have imposed blanket bans or read-only access; an expert cautioned that bans could reduce digital banking capabilities for customers.
People
Cal Al-DhubaibRubrik cloud and AI technologist who said many organizations are unprepared for AI systems to interact with their environments and security infrastructure.Chris TaleviDomino Data Lab chief information security officer who discussed the difficulty of identifying who authorized an agent and holding it accountable for its actions.Snehal FulzeleUptiq AI founder and CEO who said banks that prohibit agents from assisting with transactions risk alienating customers, and that AI agents still require human oversight.
Vendors
Domino Data Labhold an agent accountable for its actions," said Chris Talevi, chief information security officer at Domino Data Lab.Metalike Claude Code, Codex and OpenClaw that read through emails and complete tasks. In the past few months, Meta and OpenAI released consumer-facing AI agents in Muse and Dots that people can personalize and set upOpenAICode, Codex and OpenClaw that read through emails and complete tasks. In the past few months, Meta and OpenAI released consumer-facing AI agents in Muse and Dots that people can personalize and set up appointmentsRubrikRubrik cloud and AI technologist Cal Al-Dhubaib told ISMG that many organizations aren't prepared for their environments and security infrastructure to interact with AI systems.Uptiq AISnehal Fulzele, founder and CEO of Uptiq AI, said in an email that if banks continue to block agents, they risk alienating customers.
Products
Claude Codeown agents for automated workflows. Their employees have begun building agents through services like Claude Code, Codex and OpenClaw that read through emails and complete tasks. In the past few months, Meta andCodexfor automated workflows. Their employees have begun building agents through services like Claude Code, Codex and OpenClaw that read through emails and complete tasks. In the past few months, Meta and OpenAI releasedDotsand complete tasks. In the past few months, Meta and OpenAI released consumer-facing AI agents in Muse and Dots that people can personalize and set up appointments or flag potential sales.Museemails and complete tasks. In the past few months, Meta and OpenAI released consumer-facing AI agents in Muse and Dots that people can personalize and set up appointments or flag potential sales.OpenClawworkflows. Their employees have begun building agents through services like Claude Code, Codex and OpenClaw that read through emails and complete tasks. In the past few months, Meta and OpenAI released