AI Agents In Customer Bank Accounts
Banks lock out AI agents. Customers are letting them in.
‘Never share your password’ was the first rule of online banking for a long time. Now, customers are voluntarily sharing theirs with AI, according to an Axios report indicating that third-party AI agents are logging into individual accounts at major US banks using customers’ credentials. Even more concerning is that some of these agents appear to act beyond their instructions.
This comes as US financial institutions are taking action to shut agents out: JPMorgan called outside agents too risky for clients, Charles Schwab is blocking unauthorized agents, and both Citi and Bank of America are restricting access. Across the pond, the UK FCA expects personal agents to reach UK consumers soon. Against this background, financial risk leaders need a plan for how agents can be managed – before their customers make one for them.
Why are banks locking out AI agents?
Bank security assumes that a human is the one logging in, so a password gives an agent everything the customer can do – from checking balances to moving money – with no way to limit it to the task it was given. Since a customer, an approved agent and a rogue one look identical from the bank’s side, it will be unable to distinguish a helpful errand from an account takeover until the customer themselves raises a concern. To avoid this, Charles Schwab says it already blocks agents every week and expects the number will continue to grow.
Then there is liability; JPMorgan has warned that customers who share credentials carry the risk if the agent misbehaves. Asked twice whether customers could recover money moved by an unauthorized agent, US Comptroller of the Currency Jonathan Gould gave no direct answer. Customers, banks and agent providers each assume someone else will pay.
Agents can also be turned against their users through prompt injection: hiding commands in content an AI tool reads. In April, security firm Blue41 showed how hidden text in a €0.02 payment tricked a European bank’s assistant into showing a fake login request. While in this case the bank fixed the flaw, the UK’s National Cyber Security Centre warns prompt injection risk may never be fully mitigated.
Finally, there is funding risk. JPMorgan, Bank of America and Wells Fargo collectively hold $1.6 trillion in interest-free deposits, which agents could move to higher-paying accounts. Verdantix has previously warned that agents reacting to the same signal can amplify shocks (see Verdantix If AI Can Outsmart Its Creators, Can Financial Markets Stay Safe?). If millions chase the same rate at once, deposits could be withdrawn faster than liquidity models assume.
What are regulators doing?
The UK FCA’s October 19 review of AI in financial services holds organizations responsible for their own AI, but is silent on customer-side agents. HM Treasury, meanwhile, has consulted on consent, authentication and liability for AI agent payments, proposing a major overhaul of UK payments regulation. According to a survey published by the Bank of England and the FCA in October, 80% of financial sector firms run AI agents, but the risk from customers agents could be far greater.
Ironically, UK and EU open banking rules were designed partly to end ‘screen scraping’, where apps logged in with customers’ passwords. Agents are now bringing it back.
How should financial risk leaders respond?
To mitigate potential threats, financial risk leaders should prioritize:
- Giving agents an approved way in, so they no longer need customers’ passwords. Each agent should have its own login and view-only access by default, and banks should keep a register of every outside agent linked to customer accounts.
- Writing liability rules before the first loss, updating terms, refunds and fraud playbooks for agents exceeding their mandate. Only 49% of firms surveyed by the Bank of England and the FCA have AI incident playbooks.
- Stress-testing liquidity for synchronized, agent-driven deposit outflows.
Requiring a human to approves any money an agent moves can also help, but the FCA warns against “excessive reliance on human-in-the-loop controls”. With 11 million UK adults likely to use AI agents – according to the FCA's The Mills Review – blocking agents is a holding position, not a strategy.
For more risk management content, check out Verdantix insights.
About The Author

Kieron Redpath
Analyst



