Your AI Assistant Has Your Bank Password. Who Can Prove What You Authorized?

If you’re considering connecting an AI assistant to a financial account, don’t share your banking username or password. An official account connection is a different arrangement, but you should still understand exactly what access it permits.

Useful resources: Regulation E, 12 CFR § 205.6 | Submit a complaint to the CFPB | Report fraud to the FTC | Report cybercrime to the FBI’s IC3

Major U.S. banks are restricting third-party AI agents from operating inside customer accounts and making transactions, according to reporting by Axios and Reuters, published by Channel News Asia. Some customers reportedly shared their login credentials with AI assistants.

Reuters quoted a person in the banking industry saying transactions cannot be deemed unauthorized if the customer shared their credentials. That is a reported industry position, not a legal ruling that settles every situation.

The deeper issue is not only what an agent did. It is whether anyone can show what the customer actually authorized.

What’s Happening

Axios reported that some AI agents appeared to take actions independent of user instructions. Reuters reported that major banks were blocking third-party agents from operating inside accounts and making transactions. In the same Reuters report, a person familiar with the matter said Morgan Stanley had blocked third-party agents but had not seen unauthorized or rogue agents accessing its platforms.

JPMorgan has publicly described its own restrictions. CEO Jamie Dimon told CNBC the bank would not allow just any outside agent to access its systems, citing risks to clients and the bank. A JPMorgan spokesperson told CNBC that the bank does not permit third parties to obtain customer credentials and log in on retail customers’ behalf.

The spokesperson also described customer-authorized data sharing through open-banking providers. Some connections can make data, such as balances, available on a read-only basis, without allowing the agent to send money or pay bills. That restriction is the point: read-only access avoids transaction authority, but also limits what an agent can do. CNBC reported that Charles Schwab blocks unauthorized agents to maintain control over authentication.

These are institution-specific positions, and they can change. They do not establish that any named institution was breached or that widespread losses occurred.

Authorization Needs a Record

Payment disputes have generally begun with a basic question: Did the customer authorize this transaction? In its analysis of agentic payments, Lowenstein Sandler LLP describes how AI agents change that question: can anyone prove what the consumer authorized the agent to do?

Consider an agent told to find running shoes within a budget. It might stay under the spending limit but ignore other instructions about the product. The transaction could be within one boundary of authority while falling outside the customer’s intent. The critical issue becomes what instructions were given, what limits applied, and what the agent actually did.

A dispute may also involve very different possibilities: a legitimate agent that misunderstood instructions, a compromised agent acting on someone else’s instructions, or a criminal’s agent posing as a legitimate one. Early records may not make those explanations easy to distinguish.

Lowenstein notes that agentic systems may generate intent mandates, authority scopes, identity credentials, transaction records, and activity logs. Their existence alone does not answer the dispute. The records must be authenticated, preserved, tied to a particular transaction, and produced in a usable form.

Why “Unauthorized” Matters

Regulation E sets a liability framework for unauthorized electronic fund transfers, subject to the regulation’s conditions. Those conditions include the financial institution providing required disclosures and, when an access device is involved, that device being an accepted access device.

For a qualifying unauthorized transfer involving a lost or stolen access device, a consumer who notifies the institution within two business days of learning of the loss or theft may have liability limited to the lesser of $50 or the unauthorized transfers before notice. If notice comes later, the regulation describes a limit of the lesser of $500 or a specified calculation, which includes transfers before notice and depends on what the institution establishes would not have occurred with timely notice.

A consumer must also report an unauthorized transfer shown on a periodic statement within 60 days of the statement’s transmission to avoid liability for certain later transfers. The institution must extend the applicable time limits to a reasonable period when extenuating circumstances caused the delay. State law or an agreement that imposes less liability controls.

Those protections concern unauthorized transfers. If a customer voluntarily shares credentials, an institution may argue that the resulting activity was authorized. Whether an agent exceeded the authority granted, or whether another relevant circumstance changes the analysis, depends on the facts. The reported industry view is not a substitute for a legal determination.

This is general information, not legal or financial advice. The regulatory framework and bank policies can change. Check your institution’s current terms and consult your own attorney about a specific dispute.

What Evidence Could Clarify It?

A useful record connects the instruction to the transaction. In practice, that means preserving:

1. The instruction: What did you tell the agent, and when?
2. The scope: What actions were permitted, and what was off-limits?
3. The agent’s identity: Which system and account were used, and under whose credentials?
4. The transaction: What happened, when, in which session, and to what destination?
5. The connection: How do the instruction, permissions, agent, and transaction relate?

The relevant records may sit in different places: with a bank, on a device, with an assistant provider, or in transaction records. “The platform has logs” and “the logs can be authenticated and produced” are not the same statement.

Before You Connect an Agent

• Don’t give a tool your banking username and password. If your institution offers an official connection, find out what it allows; access to account data is not the same as permission to move money.
• Write down the scope you intend to grant. A dated note is useful for your own records, but it does not replace an institution’s authorization process.
• Check your institution’s current policy. An assistant’s capabilities depend on what the financial institution permits, not just what the assistant advertises.
• Save the setup records. Keep your instructions, permission screens, confirmation messages, and dates. The setup may help explain a later transaction.
• Report a disputed transaction to your institution promptly and in writing. Regulation E has notice timelines. Keep a copy of what you sent and when.

No record guarantees that a transaction will be reversed or funds recovered. A financial institution’s decision is outside your control.

Questions for Businesses

Lowenstein’s analysis also raises practical questions for organizations using automated agents: Is there a policy distinguishing verified agents from unverified automation? Do agreements address access to logs and instructions? Do approval thresholds account for a non-human initiator? Are agent records covered by retention and legal-hold procedures, and can they be preserved in a form that supports authentication?

Lowenstein’s article cites INTERPOL’s March 2026 assessment and Signifyd data as broader context on AI-related fraud. Those figures do not establish whether a particular agent-initiated transaction was authorized.

What an Investigator Can, and Cannot, Do

Where lawfully directed, a private investigator can organize records a client lawfully holds into a dated sequence, distinguish documented facts from inference, research publicly available information about a service, document contradictions, and prepare a written report for a client, attorney, or financial institution.

An investigator cannot obtain private account records without lawful authority or the account holder’s cooperation, compel a provider to produce logs, subpoena records, determine legal liability, reverse a transaction, or promise an outcome. The bank and payment provider are the first calls, not an investigator. A legal dispute about authorization belongs with an attorney.

The Record Is the Point

The technology is new; the need to document authority is not. When the question is what an agent was told, what it was allowed to do, and what it did, a clear, preserved sequence can help explain the facts.

FAQ

Are AI agents really operating inside bank accounts?

Axios and Reuters reported that agents were accessing accounts and that major banks were restricting third-party access. Reuters reported that Morgan Stanley had blocked third-party agents but had not seen unauthorized or rogue agents accessing its platforms. The reports do not establish widespread confirmed losses.

If I gave an AI assistant my bank password, is a transfer it makes unauthorized?

That cannot be determined from credential-sharing alone. A person in the banking industry told Reuters that a transaction cannot be deemed unauthorized if the customer shared credentials. The Regulation E framework applies to qualifying unauthorized transfers; a specific dispute depends on its facts. Consult an attorney.

Which banks allow it and which block it?

CNBC reported that JPMorgan blocks external AI agents from operating on retail customers’ behalf, while some customer-authorized data sharing can be read-only. Reuters reported that Morgan Stanley blocked third-party agents, and quoted Charles Schwab’s CEO saying the brokerage blocks unauthorized agents. Policies can change, so check with your institution.

What records should I keep if I use an agent?

Keep the instruction, the permitted scope, the agent’s identity, the transaction details, and records linking those items together, such as permission screens and dated confirmations.

Can a private investigator get my money back or force a bank to produce logs?

No. An investigator cannot reverse a transaction, guarantee recovery, or compel a bank or platform to produce records. An investigator may help document records you lawfully hold. Contact your financial institution first; consult an attorney about legal questions.

Sources

• Axios, “Scoop: Unvetted AI agents are accessing U.S. bank accounts”, October 22, 2026. The article was not directly accessible for review; its substance is corroborated through the Reuters report published by CNA.
• Reuters via Channel News Asia, “US banks detect unauthorized AI agents in customer accounts, Axios reports”, October 22, 2026.
• CNBC, “JPMorgan Chase blocks outside AI agents … from directly accessing accounts”, October 13, 2026.
• Lowenstein Sandler LLP, “Who Said ‘Buy’? Agentic Payments and the Evidence Problem”, October 13, 2026.
• 12 CFR § 205.6, Liability of consumer for unauthorized transfers.

Confidential Consultation

Alaska Investigations Group has more than 10 years of professional investigative experience. We can help document a sequence of what was instructed and what was executed, or research a service before you connect it. We are not law enforcement, cannot compel records or subpoena, cannot reverse transactions or recover funds, do not determine liability, and promise no outcome. We cannot accept every request.

Licensed Private Investigations Agency in Anchorage, Alaska. License 1126. Serving statewide except Fairbanks.

Request a confidential, no-obligation consultation | (907) 232-4731 | mrrems@protonmail.com