Bank of America sent customers an email this week with a request that sounds obvious and is not: do not give your banking password to your AI assistant. TechCrunch's Zack Whittaker reported on October 27 that the email, titled "Protecting your finances in the AI World," tells customers never to share their username and password with another person or an outside service, "including an AI-enabled tool or agent that offers to act on your behalf." It also puts the cost of a mistake squarely on the customer: "If you do share your credentials... you're responsible for activity resulting from that access."
TechCrunch says Bank of America did not immediately respond to a request for comment, so what follows rests on the reported email and on our earlier coverage of how banks are handling agents. It is a small notice with a large implication: a top US bank telling customers directly that handing a login to a personal AI agent is their risk, not the bank's.
TL;DR: what is confirmed and what is not
| Question | Answer |
|---|---|
| What happened? | Bank of America emailed customers a warning about sharing credentials with AI tools and agents |
| Source | TechCrunch, which quotes the email; the bank has not yet commented to TechCrunch |
| Who pays if it goes wrong? | The email says the customer is responsible for activity resulting from shared credentials |
| Is it a ban? | It is a warning and a liability statement, not a described technical block |
| Is it only one bank? | No; JPMorgan and, per the FT, Citi and Bank of America have been restricting agents |
Agents that act on your behalf need rules. Agentbeam stops the dangerous actions →
Why would a bank send this email now?
Personal AI assistants that browse, click and pay have moved from demos to everyday tools. TechCrunch notes that customers are "increasingly handing over their log-ins as personal AI assistants grow more popular." The simplest way for an agent to act inside your bank account is to log in as you, and the simplest way to enable that is to give it your password.
For a bank, that is a nightmare in three ways. An agent acting as you looks identical to an account takeover in fraud systems. The bank cannot tell whether a transfer was your intent or an agent's misreading of your instruction. And the credentials now live inside a third-party product the bank has not vetted.
What are the three risks TechCrunch lists?
TechCrunch describes the risks in plain terms:
- The agent does something you did not intend. An assistant with full login rights can move money or change settings "without your knowledge."
- The bank locks you out. If the bank thinks someone is "raiding your account," automated defenses can freeze access, which is a real problem when you need the money.
- The assistant leaks your credentials. A security flaw in the AI product could expose your username and password to financially motivated hackers, giving them a way into your account.
The third risk is the least visible and arguably the worst. A password handed to an agent may be stored in logs, memory files, or a hosted session. Prompt injection, where hidden instructions on a web page redirect an agent, is a known way to extract secrets from agents that handle sensitive data.
Who is responsible if something goes wrong?
The email's most pointed line is about liability. As TechCrunch puts it, account agreements at most banks hold customers responsible for anyone they give their login to. Many consumer protections for unauthorized transactions assume you did not voluntarily share your credentials. Once you hand a password to a third party, a bank may argue the activity was authorized.
That is general context, not legal advice, and rules vary by country and account type. If you have already shared a login with an agent, change the password and review your account agreement.
Is this part of a bigger pattern?
Yes. We have followed the shift for two weeks:
- Our post on Citi, JPMorgan and Bank of America blocking AI agents sets out what Jamie Dimon said on JPMorgan's earnings call ("You don't want rogue agents") and what the Financial Times reported about Citi and Bank of America.
- A separate report covered banks detecting unauthorized AI agents on customer accounts, with a confirmed-versus-claimed breakdown.
- On the retail side, Amazon began blocking Meta's agent, as explored in Paul Graham's take on agentic commerce.
- On the sanctioned side, Amex introduced its Intent Intelligence Framework, where a customer approves an agent purchase intent and a token gates the payment credential.
TechCrunch's piece says the banking industry is "still trying to figure out how to allow customers access to their bank accounts using AI tools," including agentic payments that let you set an agent to buy something at the best price or the right time. The warning is one half of a transition: banks are closing the credential door while they build the token door.
How does this line up with what bank executives have said?
The customer email echoes the tone executives have used in public. On JPMorgan's third-quarter call, CEO Jamie Dimon said the bank will not let just any outside agent in, as Fortune reported. The difference is audience. Dimon was speaking to investors about risk to the bank. Bank of America's email speaks to the person holding the password and tells them who carries the loss.
That shift matters because it moves the debate from "should agents be allowed" to "who is accountable when they act." Banks can defend a blanket block on technical grounds. A liability notice goes further: it sets expectations before any dispute happens, which makes it easier to deny a claim later. Customers who read only the product marketing of an assistant, which promises it can handle everything, will not see that trade-off unless the bank spells it out.
It is also worth saying what the email does not do. As quoted by TechCrunch, it does not announce a new technical control, and it does not say the bank will refuse to reimburse fraud in every agent-related case. It reminds customers of terms they already accepted. Whether any bank would treat a hijacked agent differently from a customer who typed the password into a phishing page is an open question we could not answer from public sources.
A quick checklist before you connect any assistant to money
- Does the product ask for your bank username and password? If yes, stop.
- Does it offer a token, open banking link or virtual card instead? Prefer that.
- Can you set spending limits and see a log of every action?
- Does it require your approval before it sends money?
- Can you revoke its access in one step?
How can you let an agent help without handing over your password?
Safer patterns share one trait: the agent never sees your bank login.
- Use sanctioned connections. Open banking or token-based links let you grant limited, revocable access to specific data. The agent holds a scoped token, not your password.
- Keep a human in the loop for money movement. Require your approval on every payment or transfer. Our guide to agents that pay without private keys shows how credential-free payment flows are being designed.
- Use virtual or limited cards for agent purchases, with low limits, instead of your main account.
- Read-only first. If an assistant only needs to summarize spending, export statements yourself and give it the file.
- Check the agent's own security. Ask where credentials are stored, whether they are encrypted, and whether the product has had incidents. If you cannot find answers, assume the worst.
- Rotate and monitor. If you already shared a password, change it, enable multi-factor authentication, and turn on transaction alerts.
If you build or run agents yourself, the equivalent step is to constrain what they can do before they act. AgentBeam, from the explainx.ai team, is an agent security platform that stops AI agents before they take dangerous actions.
What this means for builders
If you build a personal-finance or shopping agent, the era of "just ask for the password" is ending. Expect banks to flag credential-based sessions and to treat them as the customer's liability. Design for token-based access, explicit intent approval and audit logs from day one, and tell users plainly why you never ask for their bank login. A product that must take passwords to function is a product that banks have now publicly told customers not to trust.
For consumers the message is simpler. Convenience from an AI assistant is real, but a bank password is the one credential that should stay with you. The bank's email says the quiet part: if the agent errs or is hijacked, the contract treats it as you.
What to watch next
- Whether Bank of America comments or publishes the same guidance on its website.
- Whether other large banks send similar customer notices.
- Whether banks announce sanctioned agent access that gives customers a safe alternative.
- Whether regulators say how liability should work when a customer-authorized agent errs.
We will update this post if the bank responds or other banks follow.
Related reading
- Citi, JPMorgan and Bank of America block AI agents: what is confirmed
- Banks detect unauthorized AI agents on customer accounts
- Amex Intent Intelligence Framework for AI agent purchases
- AI agent privacy promises: Meta Muse and OpenAI
- AWS AgentCore payments for agents without private keys
- Paul Graham and Amazon banning agents
- Bank of America hired 1,000 contractors to patch Mythos-exposed flaws
Details reflect TechCrunch's October 27, 2026 report and may change if the bank comments.