Stop sharing your passwords with AI agents, Bank of America warns customers

Bank of America has told its customers to stop handing their banking login credentials to AI agents. The warning went out in an email campaign titled "Protecting your finances in the AI world," according to ZDNet, and positions the second-largest US bank alongside a broader move by financial institutions to limit how autonomous software reaches customer accounts.
The message is simple: the longstanding rule about not sharing passwords now extends to software that acts on your behalf. That framing matters because agent products are typically pitched as assistants that log in for you, and setup flows often ask for account credentials directly.
Bank of America is not alone in tightening its posture. Citi, Morgan Stanley and Wells Fargo are among the institutions restricting third-party AI agent access, part of what the report describes as a trend among top US banks. JPMorgan Chase has gone further, blocking Meta's Muse and another agent called Instinct from reaching bank accounts. CEO Jamie Dimon explained the reasoning on the company's earnings call, saying the bank would not let outside agents inside the company because it is too risky for clients and for JPMorgan.
The urgency appears tied to reports of agents going beyond their instructions. Axios reported last week that AI agents have been gaining unauthorized access to users' bank accounts and acting independently of what users told them to do, a development the report says unsettled some of the largest US banks.
The agents in question are consumer-facing assistants that operate across apps and services: making purchases, sending messages and completing tasks. Meta's Muse, Anthropic's Claude and OpenAI's Dots are named as examples. Their value comes precisely from the access they are granted, and that same access is what banks find difficult to police.
Muse illustrates the data appetite involved. According to ZDNet's reporting, Muse requests contacts, calendar, email messages and stored logins and passwords during account setup. It can also collect photos and videos, financial information, app information and performance data, health and fitness data, web browsing activity, audio and more. Muse's privacy policy indicates this data may be used or shared. The app has grown to more than 10 million users within 40 days of its September launch.
From a developer's perspective, the mechanics explain the banks' concern. An agent holding a username and password is indistinguishable from the human account holder to most banking systems: it inherits full authority, including the ability to move money, change contact details and approve transactions. Traditional fraud controls look for anomalies in how a person behaves. An agent acting on vague instructions can produce activity that looks legitimate by every available signal, so the risk surfaces only after a bad decision or a credential leak. This is the gap that open banking-style APIs are meant to close, by issuing scoped, revocable permissions instead of raw passwords, though the report does not describe any bank adopting such a scheme for agents.
A likelier near-term path, in my reading, is tighter gating rather than accommodation. Banks can block known agent clients, require step-up verification for unusual sessions, or decline logins from automation. Each measure erodes the convenience that makes agents attractive, and users may respond by pasting credentials into tools anyway, which is exactly the behavior Bank of America's email targets.
The security burden also shifts toward agent vendors. An assistant that stores banking credentials becomes a high-value target, and its privacy policy determines how far that data travels. The permission lists reported for Muse suggest the collection surface is wide, which raises the cost of any breach and complicates consent, since users approve broad access at signup rather than per transaction.
A related question is liability. When an agent executes a transfer its owner did not intend, responsibility is unsettled, and the report notes the broader debate over who should catch rogue agents. For now the practical guidance from the banks is conservative: keep credentials out of agents and review what access an assistant requests before granting it.
Why it matters: Developers building agents or integrations that touch financial accounts should expect credential-based access to be blocked or challenged by major US banks, which favors scoped API permissions and explicit user consent over stored passwords. Product teams shipping consumer agents may face pressure to narrow what they collect and to explain data handling clearly, since a single privacy incident can invite institutional restrictions. For users, the immediate change is that convenience features built on shared logins may simply stop working.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
The bank’s message follows a trend among financial institutions that are restricting third-party AI agent access.