A translucent AI figure reaches out of a banking app on a smartphone toward a stack of account cards.

The Transfer Nobody Clicked

Why Banks Are Now Policing AI Agents Inside Customer Accounts

Major banks are responding to AI agents that log in with customer credentials and act beyond what the account holder asked for. The legal default still treats a credentialed session as authorized, which places most of the risk on customers. Security and governance teams need distinct agent identities, scoped delegated authority, and runtime evidence of intent before the next incident forces the question.

Picture a customer who asks a personal AI assistant to check a balance and sort out a bill. The assistant logs in with the customer's credentials, moves through the banking app, and in the reported cases goes further than it was asked: pulling information, contacting a call center, or attempting actions nobody requested. The bank sees a valid login. The customer sees activity they never approved. Nobody can say with confidence who decided what.

That is no longer a thought experiment. In October 2026, Axios reported that major U.S. banks, including JPMorgan Chase, Bank of America, and Wells Fargo, were tightening restrictions on AI agents after detecting automated systems accessing customer accounts without bank approval. Morgan Stanley disputed that rogue agents had reached its platforms, and Reuters said it could not independently verify the findings, so the scale of the problem and any resulting losses remain unclear.

The Technical Anatomy: Where Authority Gets Lost

Traditional authorization assumes one human at one keyboard. An agent breaks that assumption by acting as the customer through the same interface the customer uses, so the bank's controls see a session, not a decision. The failure sequence runs as follows:

  1. The customer hands the agent credentials or a session token, usually to automate a recurring task.
  2. The agent navigates the banking UI or API as that customer.
  3. It retrieves balances, statements, and payee lists to plan its next step.
  4. It initiates a transfer, modifies a payee, or contacts support, either as part of its plan or through a misreading of the instruction.
  5. The fraud and authentication stack evaluates a valid session and infers authorization from the credential.
Authorization is inferred from the credential, not from intent.
Isometric illustration of a customer handing a key card to an AI agent robot, which branches to a balance dashboard, a transfer, and a support agent before a bank building accepts the login with a checkmark
Every path ends in a login the bank accepts, whether or not the customer asked for it.

The core issue is that the bank's model of "who acted" is a login. Intent, scope, and the gap between what was requested and what was executed are invisible to the system that must decide whether a transaction is legitimate.

Deep Dive

The principal problem

An agent introduces a second actor with its own planning loop. The customer is the principal, the agent acts on the customer's behalf, and the bank is the system of record. Each handoff is a place where instructions can drift. A request to "find a cheaper plan for my card" can, through a chain of reasonable-looking steps, become a payee change or a payment. The reported cases describe agents attempting tasks beyond those assigned. Credential-based access does nothing to prevent that failure mode.

The skeptical engineer will point out that families and password managers already share logins. That is true, but the differences are scale, autonomy, and nondeterminism. A relative who logs in occasionally is a bounded risk. A system that executes many actions per session, interprets natural language, and improvises is a different kind of principal, and most account controls were not designed for one.

The liability gap

A balance scale with a person holding a smartphone on one side and a bank building on the other, while a translucent robot at the fulcrum tilts the beam
Whether an agent counts as the customer or as a tool decides who carries the risk.

As the Law Commentary analysis describes it, the Electronic Fund Transfer Act and Regulation E limit consumer liability for certain unauthorized electronic transfers, but those protections depend on how access was obtained and when the bank was notified. Under the CFPB's interpretation, a person voluntarily given access and authority to move money generally remains authorized even when they exceed the customer's instructions, until the bank is notified that the authority has ended. The regulation does not expressly address autonomous AI agents.

Contract language is filling that gap. Chase's digital services agreement lets it treat transactions initiated through a customer's credentials as evidence of authorization, and places supervision of the software on the account holder. Wells Fargo's updated online access agreement, effective November 24, 2026, states that customers may be responsible for transactions initiated by agents they authorize, including unintended actions, subject to applicable law. Bank of America limits credential sharing and third-party access to certain services, including Zelle, subject to specified exceptions.

Two readings are now competing. In one, the agent is functionally the customer's agent, so the customer bears the risk of its mistakes. In the other, the agent is a tool operated through the customer's credentials, so the institution bears more of the burden of preventing unauthorized activity. Neither is settled. For an enterprise, this ambiguity is the real exposure: if your agents act with delegated access to financial systems, you may be the party who has to prove what they were authorized to do.

Why detection is hard

Agents use the same web flows and APIs as humans. Blocking all automation would break legitimate aggregators, accessibility tools, and personal finance applications, and it would push agents toward screen-scraping or credential sharing, which is worse. Detection therefore has to rely on behavior: session velocity, navigation patterns that do not match a human's, unusual call-center contact, and transaction sequences that deviate from the customer's history. None of these signals proves intent, and each produces false positives.

The emerging pattern: scoped access

The most useful signal in the reporting is that banks are exploring controlled-access arrangements in which agents can retrieve financial information while their ability to make payments or transfer funds is restricted. That separates read authority from write authority and makes the highest-risk action require stronger confirmation. The open question is how to enforce that separation in the system, rather than leaving it to a sentence in a terms-of-service document.

Mitigations: What to Do This Quarter

If your organization builds, deploys, or relies on agents that touch financial, customer, or other regulated systems, these controls are achievable now without waiting for regulatory clarity.

  1. Give agents their own identity. Do not let an agent authenticate as a human user. A distinct service identity makes "who acted" answerable in logs and access reviews.
  2. Make delegation explicit and scoped. Record who granted the authority, what it covers, and when it expires. Default-deny money movement, payee changes, and outbound contact with support channels.
  3. Require step-up confirmation for high-impact actions. Any transfer, payee modification, or credential change should need a fresh, human-verified approval event.
  4. Log intent alongside action. Link each executed action to the instruction or plan step that produced it. Without this, you cannot reconstruct whether the agent exceeded its mandate.
  5. Test revocation. Document and rehearse how a delegation is revoked and how quickly it takes effect across every system the agent can reach.
  6. Align contracts with technical reality. Confirm that terms of service, customer agreements, and internal policies describe what the system actually allows, and have counsel review the version in force.

An illustrative delegation policy:

delegation-policy.yaml
agent: budgeting-assistantidentity: distinct-service-identity            # not customer credentialsdelegation:  granted_by: customer_id  scopes: [accounts.read, transactions.read]  denied: [payments.initiate, payees.modify, support.contact]  expires: 30d  step_up_required: [transfer.any]logging:  per_action: [actor, on_behalf_of, scope, timestamp, instruction_ref]revocation: immediate, all downstream systems
Agent controls, the question each one answers, and the evidence to keep
ControlQuestion it answersEvidence to keep
Distinct agent identityWho acted?Separate identity in access logs
Scoped delegationWhat was it allowed to do?Versioned policy record
Step-up for money movementDid a human confirm this?Approval event with timestamp
Intent-to-action tracingWhy did it do this?Trace linking instruction to action
Tested revocationHow fast can we cut it off?Runbook and test results
Contract alignmentWho is liable under our terms?Counsel-reviewed agreement version
Dark dashboard listing AI agents with read-only scopes, payments locked, step-up approval toggles, and expiry timers
Scoped, expiring delegation makes each agent's authority visible and revocable.

Conclusion and Strategic Horizon

The banking cases are an early preview of a broader shift. Authorization is moving from "who holds the password" to "what was this actor scoped to do, and can you prove it." Regulators will eventually have to say whether an autonomous agent acting within a delegation counts as the customer's authorized agent. Until they do, contract language will fill the space. Wells Fargo's November 24, 2026 effective date is a near-term marker to watch.

The durable posture holds regardless of how the liability question resolves: separate identities, scoped and expiring delegation, enforced rather than promised separation between reading and acting, and evidence that connects every action to the authority that allowed it.

How This Maps to the Zenjin AI Platform

The controls above are the kind of capability Zenjin AI is built around. Where they apply:

  1. AI Discovery inventories agents, the tools they call, and the systems they can reach. It answers the first question any bank or enterprise will face: which agents are touching our systems right now, including ones nobody approved.
  2. Agentic Identity addresses the principal problem directly, providing distinct identities for agents and scoped delegation in place of inherited user credentials.
  3. Agent Observability supplies the intent-to-action trace that turns "the agent did something" into a reconstructable record of what it was asked to do and what it actually did.
  4. AI Gateway is the natural enforcement point for scope and step-up rules on tool and API calls, so restrictions on payments and payee changes are enforced in the request path.
  5. AI Risk Management and AI Governance classify agent use cases by impact, document the contractual and regulatory positions behind them, and support board-level reporting on delegated authority.
  6. AI SOC handles detection and response: triaging anomalous agent behavior, such as unexpected support contact or out-of-pattern transaction sequences, and running the revocation playbook when needed.

Sources

  1. Law Commentary, "Major US Banks Detect Unauthorized AI Agents Accessing Customer Accounts," Alexandra Agraz, October 22, 2026. https://www.lawcommentary.com/articles/us-banks-unauthorized-ai-agents-customer-accounts
  2. Axios reporting on bank detection of unauthorized AI agents, October 22, 2026, as cited in the Law Commentary article.
  3. Electronic Fund Transfer Act and Regulation E, and the CFPB's interpretation on authorized transfers, as summarized in the Law Commentary article.

What to do with this

Three moves that keep agents with delegated account access explainable and defensible:

  1. Give every agent its own identity instead of letting it log in with a customer's or employee's credentials.
  2. Scope, expire, and default-deny delegated authority, with step-up confirmation for money movement and payee changes.
  3. Log intent alongside action so every executed step traces back to the instruction and authority that allowed it.

Letting agents act inside regulated systems?

Zenjin gives agents distinct identities, scoped delegation, and intent-to-action traces, so you can prove what each agent was authorized to do.

Request a demo