AI Agent Identity Verification Arrives Inside ChatGPT

Finance teams deciding whether to let AI agents touch real money can now check the one thing that has been missing: who actually authorised the request. Proof, the identity authorisation company, has released the first version of its Model Context Protocol (MCP) server inside ChatGPT, making AI agent identity verification available to the assistant's users without any work on the part of the businesses on the receiving end. The headline feature is x401, a protocol that lets a website or service demand cryptographic evidence of who stands behind an agent's request, and of exactly what that person approved.
In practice, a ChatGPT user adds the Proof app from the assistant's app directory, signs in once, and from then on can present a verified identity whenever a site asks for it. Every consequential action, a payment, a signature, a transfer, comes back to the person for approval with a biometric before it proceeds.
What is the Agent Evidentiary Gap?
The answer matters to any business that has been turning agents away rather than serving them. AI agents can already shop, sign and move money on a customer's behalf. What they cannot do, until now, is prove that a real person approved a specific action. A password, an OAuth token, an API key or a stored card all show that an agent has access. None of them shows authority. Proof calls that distance between access and authority the Agent Evidentiary Gap, and it is the reason retailers and banks have treated incoming agents as a fraud and dispute risk rather than a customer.
The gap stopped being theoretical this autumn. In September 2026, Amazon blocked Meta's Muse shopping agent from its marketplace, citing concerns that the agent did not identify itself and moved through accounts without the retailer's consent. The episode put a number on the problem every merchant now faces: an agent's arrival, on its own, proves very little.
How does x401 verify who authorised an AI agent?
x401 works by extending HTTP 401, the web's standard "unauthorised" response. Instead of simply refusing a request, a site can use x401 to state exactly what evidence an agent must supply before proceeding. The agent carries the request; the person behind it signs it. A hardware-protected key binds that person's verified legal identity to the precise terms of the action, and the business can verify the result itself, offline, without integrating with ChatGPT or calling back to Proof. Change a single term, the amount, the merchant or the time window, and the signature fails.
Consider an agent instructed to move $25,000 from a customer's bank account. Under x401, the receiving institution does not have to trust that the agent is acting in good faith. It receives cryptographic proof that a verified human authorised that specific transfer, for that amount, at that time. That is a stronger claim than "this agent holds a valid token," and it is the claim financial institutions have wanted before letting agents through the door.
Why is Proof putting this inside ChatGPT now?
Distribution. x401 went into production on 15 October 2026, and in the weeks since, developers could already connect Proof's MCP server to ChatGPT through a manual setup. This release removes that step: the Proof app now sits in ChatGPT's app directory, so the protocol reaches the assistant's users directly. For infrastructure whose value depends on both sides turning up, reach is the point.
With the Proof MCP live in ChatGPT, x401 touches each party to an agent transaction. People add the app once and approve consequential actions as they arise. Businesses that speak x401 can expect ChatGPT users to arrive with a verifiable answer to "who sent you, and did they say yes?", and can check the signed response for themselves.
Developers keep the same protocol, specification and MCP server released in October; nothing they built changes, and their x401 challenge now reaches ChatGPT users with no setup required on the user's side.
The move fits a wider pattern. Proof demonstrated x401 at Money20/20 in Las Vegas in October, pairing it with agent payment rails so that identity, who is behind the request, and payment, how the agent settles, are handled by separate, complementary protocols.
What does Proof bring to agent transactions?
A decade in regulated identity. Proof began life as Notarize in 2015 and created the online notarisation category before rebranding. It operates a certificate authority and issues verified digital credentials only after identity proofing to the NIST IAL2 standard for remote identity verification, certified through the Kantara Initiative. Proof says it has secured more than $1 trillion in transactions for over 8,000 businesses and helped change more than 40 state laws along the way.
Founder and chief executive Pat Kinsel has argued that the real constraint on agent commerce is trust: a business needs evidence that a real person authorised an agent before it will act on the agent's instruction. By his account, ChatGPT is where many people will first ask an agent to do something that carries weight, pay a bill, sign a form, move money, which makes verifiable authorisation at that moment the thing that matters.
Why This Matters to FinanceX Readers
For banks, payment providers and merchants, the significance lies in how the identity-and-authorisation layer is being delivered: as open, issuer-neutral infrastructure that a business can verify on its own, rather than a proprietary credential tied to one card network or platform. Much of the recent agentic-commerce activity has concentrated control of the trust layer inside the large networks. x401 points the other way, toward verification that any site can demand and any issuer can satisfy.
Two caveats belong on any reader's dashboard before this reshapes a risk model. First, the assurance is only as strong as the identity proofing behind it, which is why the NIST IAL2 and certificate-authority detail carries more weight than the ChatGPT headline. Second, adoption cuts both ways: x401 only helps if the businesses on the receiving end choose to ask for it, and most have not yet. Whether agents get to do more while customers stay in control now depends less on the protocol and more on how many relying parties decide to speak it.



