Your AI Agent Wants Your Bank Login. Who Pays If It Errs?

On September 22, Dana gave a new AI agent her bank login and asked it to pay her $186.40 electric bill. The agent paid $1,864.00. Her checking account went negative two days before rent was due, and when she called her bank to report an unauthorized transfer, the rep asked a simple question: did you give this service your username and password? (Dana is a composite, and I'm not describing a reported bug in any specific product. The bank's question is the real part.)

If you've been tempted by the wave of AI agents that promise to pay bills, cancel subscriptions and hunt down savings for you, that question is the one to think about before you hand over a bank login. The fraud protections most of us count on for our checking accounts were written around a simple idea: someone you didn't authorize took your money. An agent you set up yourself doesn't fit that idea neatly, and nobody has tested in court where the line falls.

What these agents can do now

Meta introduced Muse on September 8, 2026, for U.S. users on the web, in iOS and Android apps, and inside WhatsApp, according to TechCrunch's launch coverage. It connects to your email, calendar, payment and shopping apps, and Meta pitches it for tasks like lowering bills, filling out forms and making purchases. The basic tier is free, with paid plans at $20 and $100 a month for heavier use.

The part that matters for your bank account is how it connects. TechCrunch reported that Muse ships with built-in connectors, lets you supply credentials for services with a public API, and can fall back on running a web browser for sites that don't have one. Meta says credentials you share go into secure storage so the agent can use them without seeing them. In practice, the agent logs in as you.

Not every company is comfortable with that. By late September, Amazon had blocked Muse from its site over terms-of-service concerns, Android Headlines reported on September 28. Muse is the splashiest launch, but it has company. OpenAI and others have offered agents that drive a web browser since 2025, and the card networks have been preparing for agent shopping since Visa announced its Intelligent Commerce platform on April 30, 2025.

Demand looks real, too. A 2025 Boston Consulting Group study, cited in a January 2026 white paper from the Consumer Bankers Association, found that 42% of consumers would let AI shop entirely on their behalf in at least one product category.

The Regulation E rule nobody mentions in the demo

Your checking account and debit card are covered by the Electronic Fund Transfer Act and its rule, Regulation E. If a thief drains your account, your losses are capped. The CFPB's text of the rule limits you to $50 if you report within two business days of learning your card or login was compromised, and up to $500 if you wait longer. Miss the 60-day window after your statement goes out and you can be on the hook for transfers after that point.

Those caps only apply to unauthorized transfers. And the official interpretation of Regulation E, comment 2(m)-2, says this: if you give your access device to someone and grant them authority to make transfers, and that person goes beyond what you allowed, "the consumer is fully liable for the transfers" until you tell your bank that person is no longer authorized.

The rule was written with a family member or coworker in mind. Swap in an AI agent and you can see the problem. You gave it the credentials. You told it to pay bills. When it sends ten times the right amount, your bank has a reasonable argument that the transfer was authorized and that you're fully liable.

The Consumer Bankers Association flagged this exact gap in its January 2026 white paper on agentic payments. It listed as an open question whether authorizing an AI agent to use an account counts as providing an access device, and what happens when the agent exceeds its authority or gets compromised. There's a separate comment in the rule saying a device obtained through fraud or robbery is treated as unauthorized, so a hacked agent might land on your side of the line. I'd lean hard on that "might." No regulator or court has ruled on it, and I wouldn't bet my rent on how the first case goes.

Two more wrinkles make it worse. Regulation E doesn't cover wire transfers at all, so if an agent can reach your bank's wire screen, you're relying on contract law and your bank's goodwill. And many online banking agreements tell you not to share your password with anyone. Breaking that term won't erase your federal rights, but it gives the bank one more reason to push back on a claim. Pull up your own bank's online access agreement and search it for "password" and "third party" before you link anything.

Read-only access versus handing over your login

A lot of the useful stuff an agent does for your money only requires reading. Spotting a subscription you forgot or a savings account paying next to nothing works fine from your transaction history.

Read-only access usually runs through an aggregator like Plaid, Finicity or Akoya. You log in on your bank's own page, the bank hands the app a token with limited permissions, and the app never holds your password. You can revoke that token from your bank's app or website without changing your login. Handing an agent your actual username and password is a different arrangement. It can see everything you can see and do everything you can do, including moving money.

The rules that were supposed to standardize this are stuck. The CFPB finalized its open banking rule under Section 1033 in October 2024, but a federal court in the Eastern District of Kentucky blocked enforcement, and the first compliance date of April 1, 2026 came and went without binding anyone, according to an April 9, 2026 analysis by the law firm Cozen O'Connor. The bureau has been reconsidering the rule since August 2025. For now, whether an app uses safe tokens or your raw password depends on the app and your bank.

How to let an agent help without handing over the vault

Start with read-only. Connect the agent through your bank's aggregator flow and let it find the problems: the duplicate streaming plan, the $15 monthly maintenance fee, the savings account paying next to nothing. Then you click the cancel button or make the transfer yourself. You lose a minute of convenience and keep every one of your federal protections.

For anything that involves spending, put a credit card in front of the agent, ideally a virtual card number with a spending limit. Credit cards fall under a different law. Regulation Z caps your liability for unauthorized use at $50, and the Fair Credit Billing Act lets you dispute billing errors like duplicate charges or goods that never arrived. When a charge goes wrong on a credit card, the issuer's money is at stake while the dispute is open. When it goes wrong on a debit card or a bank transfer, yours is. Visa's agent platform was built around this model, with tokenized card credentials and spending limits you set.

If you want an agent paying bills straight from a bank account, give it a separate checking account that holds only what it needs. Open a free second account, move in next month's bill money, and link the agent there. If it makes Dana's mistake, the damage stops at that balance and your rent money sits somewhere it can't touch. Turn on text alerts for every transaction on that account, since you'll want to catch a bad payment the same day.

Write down every account you've connected and how. When you stop using an agent, or it starts doing things you didn't ask for, change that password and send your bank a short written message saying the service is no longer authorized to make transfers. Comment 2(m)-2 ties your liability to that notice, so dated proof matters. An email or secure message through your bank's app works better than a phone call nobody logged.

If an agent makes a bad payment

Move fast, because the Regulation E clock starts when you learn about the problem, and the $50 cap needs notice within two business days. Call your bank, then follow up in writing the same day. Describe exactly what you told the agent to do and what it did instead. "I authorized a $186.40 payment and $1,864.00 was sent" is a stronger claim than "I didn't authorize this," and it's honest.

Contact the payee too. A utility or landlord that received an overpayment will often refund it or credit your account, and that can be quicker than any bank dispute. Check the agent company's terms for an error or refund process. If your bank denies the claim and you think it's wrong, the CFPB still takes complaints at consumerfinance.gov.

Bottom Line

AI agents can save you real money on the boring stuff, and I expect they'll get better at it fast. The law hasn't caught up, though, and your bank login is the one thing I'd hold back until it does. Here's what I'd do this week:

1. Open your bank's online access agreement and search it for "password" and "third party" so you know what it says about sharing credentials.
2. If you've already given any app or agent your bank username and password, change the password today and reconnect it through a read-only aggregator link instead.
3. Set up a virtual card number with a monthly limit for any agent that shops or pays for things on your behalf.
4. Turn on real-time transaction alerts for every account an agent can touch, so a $1,864 mistake shows up on your phone the same day.