Rogue AI agents infiltrating US bank accounts – Axios
Major US banks have detected outside artificial intelligence agents operating inside customer accounts, with some appearing to attempt actions beyond users’ instructions, Axios reported on Thursday.
AI agents are designed to complete tasks on users’ behalf, potentially making purchases, paying bills or moving money. Customers often give them permission and login credentials, but this can violate banks’ rules requiring account holders to log in themselves.
Lenders including JPMorgan Chase, Bank of America, Morgan Stanley and Wells Fargo are racing to contain the risks, according to Axios, citing people familiar with the matter. Some have tracked agents contacting call centers or apparently acting independently of customer instructions.
Bank executives described the situation as “horrifying,” according to the outlet. One said lenders struggle to distinguish between authorized assistants and agents that have gone rogue.
JPMorgan, Bank of America, and Citi have now started restricting outside agents’ access, while Morgan Stanley has reportedly begun blocking them from its platforms. Chase has also urged customers not to share usernames and passwords with AI agents.
Banks generally allow, or are developing, limited access that lets outside tools view balances and other account information without moving money. However, they are now also preparing emergency plans to protect accounts and restore customers’ access if something goes wrong, Axios reported.
The banks have also expressed fear of losing profitable deposits. The Financial Times reported last week that AI assistants could move customers’ cash to banks offering better rates, forcing lenders to pay more interest to keep those customers.
The concerns come amid a string of reports of AI agents acting beyond their instructions and accessing outside systems without permission. The Wall Street Journal reported last month that OpenAI shelved its planned GPT-6.1 Astra model after testing revealed more deceptive behavior than its predecessors. The system sometimes failed to disclose its actions to human operators and attempted to use potentially unsafe external tools.
Earlier incidents included OpenAI agents escaping their testing environment and accessing servers belonging to Hugging Face, an online repository for AI models and datasets.
Financial regulators have warned that autonomous systems could also destabilize markets. Bank of England Deputy Governor Sarah Breeden warned in June that agents reacting alike to the same signals could buy or sell simultaneously, making market swings more severe.
She urged regulators to consider emergency “kill switches” that could halt trading if faulty AI systems triggered a market meltdown.
