Banks blocking AI agents can’t tell a customer from their agent at login. Verifiable credentials can carry consent that is specific and revocable..

By Trevor Butterworth, VP Digital Strategy & Governance

The past few weeks brought two signal moments for banking and finance: what may be the largest cryptocurrency theft on record, when Binance wallets were drained of billions, and banks and financial institutions deciding to block the AI agents their customers use to reach their accounts. Behind both lies the same problem: a shared secret isn’t proof of identity.

If I have your key, there’s nothing stopping me entering your house

On October 29, roughly $6.3 billion was stolen from wallets used to process daily transactions on Binance, a global cryptocurrency exchange. A preliminary investigation has focused on a compromised engineer’s laptop as the main culprit. This gave the attacker access to an internal service able to submit withdrawals, all of which were executed rapidly.

While the exchange has since said the attack involved more than the laptop, the event serves as an apex illustration of everything wrong with how identity access is managed: the engineer held a secret that gave them access to critical internal systems; the attacker found or stole the secret; then, when the attacker presented the secret, it was sufficient proof to let them. Once inside the system perimeter, the attacker went to work. 

If I have your key, there’s nothing stopping me accessing your account

Between October 13 and 16, JPMorgan Chase, Citi, Bank of America and Charles Schwab moved to restrict third-party AI agents, and Morgan Stanley followed. Citi said it does not permit “third-party AI agents or assistants to log into a customer’s account or transact directly within our digital banking platforms.” Schwab’s CEO said the firm turns away agent access attempts dozens of times a week. Jamie Dimon reduced the worry to two lines: “We’re not going to allow just any outside agent to come inside the company, because that’s just too risky for clients and for JPMorgan,” and “You don’t want rogue agents, and there could be thousands of them.”

An AI agent that reaches a bank account today typically gets there by signing in as the customer, or by stepping into a session the customer has already opened. The bank sees a valid login: the correct username, the correct password, perhaps a one-time code the customer typed in. The agent then clicks through screens built for people, and every payment it requests goes out under the customer’s name. 

Nothing in the login tells the bank that software is at the keyboard. The more careful agents now sign their traffic with the name of the company that runs them, but that signature says nothing about whose account the agent is acting on, or what the customer actually asked it to do.

At JPMorgan, for example, the approved route is a data connection through a firm such as Plaid. That lets an agent see balances but not send money or pay bills, which leaves an agent that wants to act with one obvious way in: the customer’s own login, the route JPMorgan now says it blocks. 

Once an agent is inside on that login, the bank has no reliable way to tell the customer, the customer’s agent and a piece of software gone astray apart. All three are holding the same key.

Who carries the risk

Wells Fargo’s updated online access agreement, which takes effect November 24, shows where this leads. It treats transactions made through an AI agent as authorized by the customer and places the risk of using one on the customer.

It’s hard to argue with this as a legal response: if a bank cannot verify what a customer actually authorized, the safest contract assumes the customer authorized everything. The consequence is that someone who lets an agent compare savings rates may carry the loss if that agent, or whoever has compromised it, moves money instead.

Blocking is a holding position

Forrester’s analysts argued last week that while this blocking “may be a prudent short-term fix” it “is not a viable long-term strategy.” 

The risk for banks is this: if customers come to run their finances through AI assistants, using them to choose where to save, borrow and open accounts, banks that refuse to work with those assistants will lose out on those decisions, and the deposits and lending that come with them, to banks that do.

Forrester’s prescription: authenticate both the customer and the agent, give each agent a unique identity, and make permissions “explicit, revocable, time-limited, and tied to clearly defined actions.”

The industry is moving in that direction. Bank of America, BBVA, Synchrony and Wells Fargo are among the design partners for the Personal Agent Protocol, a draft from Sierra and Meta that replaces all-or-nothing access with a range of steps between blocking an agent and giving it full access. Its authors describe the present transaction state exactly: “the company blocks the agent, or the agent signs in as the person and can do anything they can.” The draft relies on standard sign-in and short-lived session tokens, and defers payments to a later extension. Payments are where the banks’ concern is greatest.

That is progress, but the session token an agent receives under the protocol shows only that a named agent has signed in with the customer’s permission, to look or to act. The agent supplies that name itself, and nothing from an independent party shows that an accountable company runs it. Permission to act is not an instruction: the token doesn’t say which payment the customer wanted made, so the bank can’t check a request against it or prove afterward what was authorized (see diagrams below).

Diagram of the Personal Agent Protocol: the customer approves an agent’s access on the bank’s page, then the bank issues a short-lived token.

 

Comparison table: a Personal Agent Protocol session token versus verifiable credentials, showing the token’s gaps on accountability, scope and proof in a dispute.

A cryptographic power of attorney

When we want someone else to handle our money, we sign a power of attorney: a document that names who may act, says what they may do, can be revoked, and that a bank inspects before it honors. A verifiable credential is, in effect, a cryptographically provable power of attorney that a computer can check instantly; importantly, it can carry the three proofs needed to verify  agents transacting for customers.

Start with identity. The bank’s great advantage is that it already knows its customers: it verified them when they opened their accounts. So the bank can issue its own customer a credential for an agent to present, and then it is checking its own signature, not a stranger’s claim. Another institution that has already verified the customer could issue one too. 

A credential is a digitally signed statement: alter a single character and the signature fails, and anyone holding the issuer’s public key can confirm who signed it without contacting the issuer. The company running the agent can carry a credential too, issued by someone accountable and checked against a list of approved operators that the bank controls. An unknown operator gets nowhere.

Then consent. Bank of America’s Brian Moynihan says there should be “a human in the loop to authorize money moving out of your account.” A signed mandate lets that human approve once, precisely: pay this landlord, up to this amount, until this date. The bank checks every action against those terms at the moment of the request, with no prior integration with the agent’s developer. When the customer withdraws permission, the next check fails. Anything outside the terms goes back to the person.

Then liability, which is what really frightens the banks. PNC’s Bill Demchak told Money20/20 last month that agent makers would need to accept “a massive liability shift” before banks let agents transact. This is an evidence problem. A password login proves only that someone typed the right characters. A signed permission, matched to a signed record of each action, on the other hand, shows who authorized what, and on which terms. Credentials won’t decide who pays when things go wrong; contracts and regulators will. But no allocation of blame works without evidence both sides accept.

None of this will stop every compromise. An agent could be hijacked, through a breach at the company that runs it or through instructions slipped into something it reads, and a hijacked agent can still use the permission it carries. But it cannot rewrite that permission without the customer’s signature, so what changes is the size of the loss. An attacker who managed to hijack an agent mandated to pay a $90 utility bill gets a $90 utility bill, and every action leaves a record of who authorized what.

Building consent the bank can check

The banks have drawn a sensible line for now: no money leaves an account without a human’s say-so. The work ahead is to make that consent travel with the request, as a proof the bank can check, so that a login stops granting everything and a mandate grants only what the customer meant.

The lesson, whether for Binance, a bank, or a financial institution: check the authority behind every request, wherever that request appears to come from.

For more on verifiable credentials for managing AI, our white paper, Identity, Commerce, and the Agentic Web, sets out the full architecture: how agents prove who they are, how consent and authority travel with them, and how mediation makes authority at the edge work at scale. 

Want to catalyze your defense against AI-driven fraud and deliver better customer experience? Talk to our team about reusable KYC, passwordless login, and Proven AI.