A Credential Is Not a Mandate

Five of the largest financial institutions in the United States found third-party AI agents inside customer accounts in October and did the same thing. They shut the door. One bank executive named the real problem out loud.

In October, five of the largest financial institutions in the United States found third-party AI agents inside customer accounts. Every one of them responded the same way. They shut the door.

Citigroup refuses third-party agent logins and transactions outright, allowing authorized open-banking providers a read-only view. Charles Schwab blocks unauthorized agents, including Meta's Muse, at a rate its chief executive put at the tens and twenties each week, and points clients toward its own assistant instead. JPMorgan lets an outside agent read a balance through a connector and move no money, with no timeline for changing that.

One bank executive told Axios what the actual problem was. They cannot tell whether an agent was authorized by the customer or has gone rogue.

That sentence is the whole story. The bank can see a valid credential. It cannot see a mandate.

Thirty years of answering the wrong question

Access control has answered one question since the 1990s. Who are you. That was sufficient for exactly as long as the thing holding the credential was also the thing deciding what to do with it. A password belonged to a person, and the person arrived with intent. Proving identity was a reasonable proxy for establishing authority, because the two were never separated.

They are separated now. Software holds the credential and forms its own intentions about how to use it. The proxy broke, and almost every control built on top of it inherited the break.

The industry has also run this exact play before. Screen scraping arrived, banks blocked it, and a decade of standards work produced open banking: a governed channel with consent, defined scope, and revocation. That channel governs reads. Nobody built the write-side equivalent. So the first time software began moving money on a customer's behalf, the industry reached for the only instrument it had in 2009. A locked door.

A bank wrote the specification

Two things happened in the same month, from opposite directions.

The banks blocked. And on September 29, Zack Anderson, Chief Data and Analytics Officer for Payments and Global Banking at J.P. Morgan, published the architecture that would make blocking unnecessary.

The paper is titled A prompt is not a control. It argues for an independent policy engine that checks the agent, who it represents, and its mandate. It argues that exposure must be counted across every agent acting under the same mandate, including transactions still in flight. It argues that the execution service has to enforce the decision before cash moves, through a path the agent cannot bypass or reconfigure. It argues that the decision and the policy version stay bound to the transaction, in a record the agents themselves cannot alter. And it argues that a message from one agent must never, on its own, widen another agent's mandate.

Set those requirements beside the five blocking announcements. A bank has written the specification for the control it is currently substituting with a lock.

A credential is not a mandate

A credential answers who is acting. A mandate answers what that actor was permitted to do, by whom, within what limit, and until when. Authentication produces the first. Nothing in the stack today produces the second.

This is why detection keeps arriving too late to matter. Arkose Labs found that 2.3 percent of legitimate agent sessions carried a verified agent identity signature in the third quarter, across every industry it monitors. Improving that number is worth doing, and it would still only tell a bank which agent is at the door. It would say nothing about what the customer agreed to.

A mandate also cannot live inside one institution. A limit a person sets reaches across every account they hold, at every institution, through every agent acting on their behalf. No single bank can see that. No single bank should hold the master copy of it either, because the last thirty years of data governance is a long record of what happens when one party volunteers to be the master. The mandate has to be resolved in flight, at the moment of the transaction, from the authorities that already hold the pieces.

Federation, not mastery. The argument has not changed. The stakes have.

What this requires of each party

For banks, blocking buys time and does not buy safety. A customer who wants an agent will find a worse channel to use one through, and the institution loses visibility rather than gaining it. Blocking also leaves the liability question untouched, which is why Wells Fargo will hold customers liable for agent activity after November 24 unless they can prove they revoked the agent's access, and why Bank of America is limiting deposit access while it writes protocols to decide which party pays.

Read the Wells Fargo rule carefully, because it names the real gap. A customer must prove a revocation. No person can prove a revocation they hold no record of. Whoever holds the evidence decides the dispute, and at present that is nobody.

For agent operators, five institutions have closed and assurances will not reopen them. What reopens them is arriving with a mandate rather than a credential, and leaving behind a decision record anyone can verify without trusting the party who produced it.

For regulators, the question is no longer whether agents will transact. It is who carries the burden of proof when one transacts wrongly. That burden is settling, quietly and by default, on the individual least able to discharge it.

Until something can say what the customer agreed to, every institution will keep choosing the lock.

A credential proves who logged in. It cannot say what they agreed to.

Cross-link: this piece is the arc's first essay to engage a specific named published architecture from a major financial institution and to find it in substantive agreement with the arc. Zack Anderson's paper "A prompt is not a control: What do recent industry incidents tell us about agentic controls?" was published by J.P. Morgan Payments on September 29, 2026 (jpmorgan.com/payments/newsroom/ai-agent-controls-payments), and specifies five requirements that map directly onto the architecture named across N° 020 (The Authorization Gap), N° 029 (Detection Is Downstream of Authorization), N° 030 (A Signature Is Not a Premise), N° 031 (Card Fraud Defense Rests on One Assumption), and N° 033 (A Governor You Can Import Is Not a Boundary): independent policy engine, exposure counted across every agent under one mandate, execution-side enforcement before cash moves on a path the agent cannot bypass, decision-and-policy-version bound to the transaction in a record the agents cannot alter, and no cross-agent mandate widening. The five-bank blocking actions in October 2026 are verified against reporting in ibtimes.co.uk, Financial Times, Axios, and American Banker: JPMorgan, Citigroup, Bank of America, Morgan Stanley, and Charles Schwab. Charles Schwab CEO Rick Wurster's "tens and twenties each week" rate is from his statement to investmentnews.com. Arkose Labs' finding that 2.3 percent of legitimate agent sessions carry a verified agent identity signature is from its Q3 2026 Agentic AI Trust Report. Wells Fargo's November 24, 2026 liability shift is confirmed; the specific framing "unless they can prove they revoked the agent's access" is this essay's reading of the practical burden the rule places on customers, since the published rule shifts liability where credentials have been shared with an agent and provides no reciprocal mechanism for a customer to demonstrate revocation. Bank of America CEO Brian Moynihan has publicly argued (October 14, 2026) that a human must remain involved before money leaves an account. The federation-over-mastery lineage is developed most fully in N° 004 (Federation, Not Mastery), N° 011 (The Control Plane Cannot Be Where the Data Sits), and N° 034 (Coordination Is Not Authorization). Reciprocal: the control plane for the mandate decision lives at federatediq.ai.

End N° 037