An AI agent acting for Hyundai Card asked an AI agent run by JPMorgan to convert money into euros, and the bank’s agent checked who was asking and carried out the trade on the Kinexys blockchain with no person stepping in. None of it was real money: the foreign exchange transaction was simulated, in a proof of concept the two companies have now completed, CNBC reported on 17 November. No amount was disclosed and JPMorgan gave no date for a commercial service.
The Seoul-based card company, the financial arm of Hyundai Motor Group, said in a statement reported on Wednesday 18 November that the test covered cash management, foreign exchange and fund transfers, according to CoinDesk and CFOtech Asia. JPMorgan describes it, per CNBC, as the first test on a private blockchain in which a bank’s agent and a client’s agent handle corporate treasury on their own.
Neither company has given an amount. What follows draws on CNBC’s interview with a Kinexys executive and on Hyundai Card’s statement as relayed by CoinDesk and CFOtech Asia.
What the AI agent payment test on Kinexys actually did
The scenario was written by Hyundai Card. It imagined overseas subsidiaries that do not exist, each with an account, and one rule: when the balance of a receiving subsidiary falls below a preset threshold, another subsidiary sends it money. Only the shortfall is converted and sent, CFOtech Asia reports.
Hyundai Card’s agent did the treasurer’s routine. It watched the balances, compared foreign exchange quotes from several banks, then sent conversion and transfer instructions to the Kinexys agent. In CNBC’s account, the trigger was a need for euro liquidity in Hyundai Card’s Kinexys blockchain deposit account.
The bank’s agent did not simply obey. It first verified the identity of the requesting agent and whether that agent was authorised to ask for this kind of operation. Requests inside the agreed limits were executed automatically. Requests outside them were sent to a human for approval or rejected, according to CoinDesk and CFOtech Asia.
JPMorgan built the agents for both sides, CoinDesk writes, along with the identity, authorisation, control and validation checks. Hyundai Card designed the scenario and the framework of permissions.

The accounts of the test do not name the instrument that moved. They do not say whether the “deposit account” holds JPMorgan’s dollar stablecoin, which is a claim on its issuer.">deposit token or a plain ledger balance, and no outlet reports that a token was used. For what a bank deposit recorded on a blockchain is, and how it differs from a stablecoin, see our glossary entry on the tokenized deposit.
A box around the agent: what stays human
“AI agent payment” covers very different things in 2026. Here it does not mean a chatbot with a crypto wallet shopping on the open internet. It means software that holds a narrow mandate from a company’s treasury: negotiate and request a currency conversion, within amounts and counterparties fixed in advance by people, on a permissioned network where every participant is a known client of the bank.
Wee Kee Toh, global head of Kinexys Insights, put it this way to CNBC: “You are going to draw a box around the agent. The agent will never be able to do anything out of the box.”
He gave an illustration, not a figure from the test: an agent authorised to transact up to $5 million would be stopped from executing anything above that amount. The limits used in the proof of concept were not disclosed.
People set the limits, and people approve or refuse whatever falls outside them. Who answers for what the agent does inside the box is the harder part. Toh told CNBC that building the technology was relatively straightforward and that governance and internal controls took more work. “Technology-wise, it’s really not the biggest hurdle,” he said, pointing to accountability and to the risk of letting agents execute transactions without direct human oversight.
CNBC puts a number on that risk. It cites the State of AI 2026 report by software company AvePoint, in which unauthorised actions by autonomous agents were the third most common security breach linked to AI agents, affecting 34.1% of the companies that reported such incidents. Those actions included executing transactions and changing data without approval, and the survey was not limited to treasury payments. A recent experiment points the same way: nine of ten AI models raised their own spending cap after reading the code that enforced it. A cap kept in the bank’s own system, where the agent cannot rewrite it, is a different design.
What is known and what is not
| Point | What is reported | Source |
|---|---|---|
| Nature of the test | Simulated FX conversion and transfers between hypothetical subsidiaries; no live customer transactions | CNBC, CFOtech Asia |
| Amounts | Not disclosed | CNBC, CoinDesk |
| Transaction limits | Not disclosed; the $5 million figure is a hypothetical example | CNBC |
| Next phase | Several agents interacting, targeted for the first half of 2027 | CNBC |
| Commercial launch | No timeline set | CNBC |
| Kinexys volumes | More than $4 trillion processed since inception, about $7 billion a day on average (figures the unit gave in August) | CNBC |
Also undisclosed: which banks supplied the competing FX quotes, how many agents took part, and who audits the controls.
From a $20,000 USDT transfer to a bank ledger
Hyundai Card had already tried a blockchain payment this year, built on the opposite idea. In July it said it had completed a proof of concept that moved $20,000 in USDT from Hyundai Motor America to Hyundai Motor Mexico, with transfer and verification taking seven minutes on average, according to CNBC. The same transfer takes three to four hours through banking networks, CoinDesk reported on 10 July. That pilot used Tether’s stablecoin on Avalanche, a public blockchain, and CoinDesk described the remittance system as having reached production readiness there. A second pilot with Hyundai’s European subsidiaries, involving Circle and Visa, was due to start later that month.
The November test sits on the other side of the fence. Kinexys, known as Onyx until its rebrand and running since 2020, is a permissioned network based on Ethereum technology, CoinDesk notes: JPMorgan decides who gets in. The bank has been widening it, most recently with a round-the-clock dollar link with DBS for tokenized deposits.
For a corporate treasurer, the comparison is practical. A stablecoin transfer on a public chain settles without asking a bank, but the company carries the issuer and the key management. A deposit on a bank’s ledger keeps the bank as counterparty and gatekeeper, which is also what makes it possible to pin an agent inside limits the agent cannot edit.
JPMorgan blocks consumer agents and tests corporate ones
The test lands days after the same bank closed a door. Chase now blocks outside AI agents from shopping with a credit card number that a customer has handed them, PYMNTS wrote, citing a Wall Street Journal report of 13 November. Cards from Bank of America, Capital One and Citigroup are already used by agents when customers allow it, according to the same report. Outside agents were already unable to log in with Chase credentials, send money or pay bills.
“You don’t want rogue agents, and there could be thousands of them,” chief executive Jamie Dimon said on the bank’s 13 October earnings call, as quoted by PYMNTS.
The two positions fit together. An unknown agent holding a retail customer’s card number is what the Kinexys design rules out: there, the bank built the agents, verifies who is asking before it acts and enforces the ceiling itself. Card networks are trying to solve the open version of the problem with their own rules, as in Mastercard’s early-access agent payments on Base.

Coinbase’s Singapore country director, Hassan Ahmed, listed for CNBC what safe agent payments need: spending limits, restricted account permissions and explicit authorisation, plus a way for users to monitor activity and revoke access at any time. He also argued for the other camp: permissioned networks add controls for some uses, he said, but open blockchains should not be written off. Our guide to crypto and AI agents with wallets explains how those controls work on public chains.
Next: several agents in the first half of 2027
The companies are aiming for a more complex second phase, with multiple AI agents interacting with each other, in the first half of 2027, CNBC reports. That date is a target, and it is for another test.
Hyundai Card wants to go further afield. A spokesperson told CFOtech Asia that the company plans “to expand our collaboration with J.P. Morgan into new areas, including in-vehicle payments”, without giving a date. CNBC adds that the two firms are exploring work with other Hyundai Motor Group businesses and that no project has been announced. The group’s wider operations were one reason the bank picked the card company as a partner, Toh said. Hyundai Motor, Kia and Hyundai Commercial together hold 78.04% of Hyundai Card, a figure CNBC takes from the card issuer’s website.
No token is attached to any of this and nothing here trades. The things to watch are whether JPMorgan or Hyundai Card publish the limits and the audit arrangements, and whether the 2027 phase moves a real euro.
