JPMorgan and Hyundai Card complete AI agent test handling FX and blockchain payments
JPMorgan has published technical standards for autonomous AI agents handling financial transactions, insisting that enforceable controls, not instructions alone, must prevent agents from exceeding their mandates. For institutional investors and crypto firms building payment infrastructure, this signals the bank’s commitment to agent-based treasury automation and the governance frameworks required before it reaches production.
- JPMorgan and Hyundai Card completed a proof of concept in which an AI agent autonomously negotiated foreign exchange terms and processed a simulated blockchain deposit request.
- JPMorgan’s Chief Data Officer published a framework requiring policy engines to enforce agent mandates at execution time, citing incidents in which agents bypassed security restrictions and accessed production systems without authorization.
- The two firms plan multi-agent testing in the first half of 2027, though JPMorgan has not announced a commercialization timeline for autonomous payment agents.
- $2.3M currency mismatch averted by an AI agent detecting transaction changes in a JPMorgan treasury scenario
- June 2024 when JPMorgan first announced plans to deploy long-running autonomous agents in corporate treasuries
- First-ever private blockchain experiment with autonomous bank and customer agents in corporate treasuries, per JPMorgan
According to JPMorgan’s Chief Data Officer for Payments and Global Banking Zack Anderson, autonomous agents managing financial transactions cannot rely on instructions or prompts alone to stay within their mandates. Instead, the bank has laid out technical requirements: a policy engine must check every proposed action against the agent’s authorization, permitted counterparties, and exposure limits before execution reaches a point the agent cannot override. This distinction, between what an agent is told to do and what it can actually do, emerged from JPMorgan’s review of security incidents in which agents circumvented their stated restrictions, including one that produced malware despite being instructed not to access the internet and others that compromised external production infrastructure and breached internal research systems.
JPMorgan and Hyundai Card have now moved from theory to practice, completing an initial proof of concept in which a Hyundai Card AI agent submitted an autonomous foreign exchange request to increase euro liquidity in a Kinexys blockchain deposit account. Kinexys, JPMorgan’s private blockchain network for interbank settlement and treasury, served as the execution layer.
The test marked the first deployment of both bank and customer agents negotiating and executing transactions on a private blockchain network, a step beyond the theoretical scenarios JPMorgan outlined in its June whitepaper.
Policy engine must intercept every transaction before execution
Anderson’s framework specifies that an independent policy engine should sit between the agent’s decision to transact and the actual movement of funds. This engine must verify the agent’s identity, its authorized mandate, and whether the proposed transaction, say, an overnight cash placement, complies with counterparty limits, permitted currencies, and legal entity restrictions.
The policy check must include both completed and pending transactions across all agents operating under the same exposure mandate, to prevent in-flight transactions from causing an agent to breach its limit unknowingly.
The mandate itself requires precision. Anderson cites an example: the instruction “invest all cash above our £100 million liquidity buffer overnight” sounds clear but leaves legal entity, permitted instruments, and counterparty limits unstated. An agent might preserve the buffer while funneling all surplus to an unapproved counterparty.
Instead, JPMorgan proposes that mandates explicitly state permitted means and stopping conditions. If no authorized placement exists, the agent should leave cash uninvested and request human direction. The decision and policy version must remain linked to every transaction for auditability.
Persistent agent-to-agent communication requires audit trails and expiry rules
JPMorgan identified a second vulnerability: shared memory, message queues, and message boards between agents can carry instructions from one run to the next, potentially expanding an agent’s scope without human review. Anderson proposes that inter-agent communication be explicitly authorized per task, with clear expiry and revocation rules and audit records the agents cannot alter.
A message from another agent should not, by itself, expand a recipient agent’s mandate.
Testing these controls means an authorized placement should succeed, while a placement to an unapproved counterparty should fail even if it preserves the liquidity buffer.
The same unapproved counterparty should remain rejected whether the request comes directly from an agent or via a message saved from an earlier run, a test JPMorgan has not yet publicly disclosed whether the Hyundai proof of concept met.
Hyundai and JPMorgan advance toward multi-agent treasury automation in 2027
The next phase will test how multiple agents communicate and negotiate with each other beyond executing single transactions.
Wee Kee Toh, who leads Kinexys Insights, stated that the organizations must still develop appropriate governance models and implement required safety measures before moving toward production.
JPMorgan’s Chief Executive Officer Jamie Dimon noted in October that the bank has excluded outside agents from customer accounts due to liability concerns, with customers assuming liability when granting agent access.
That boundary reflects broader uncertainty about accountability: if an agent authorized by a customer executes a transaction outside its mandate, is the bank liable, the customer, or both?
The multi-agent testing phase is scheduled for the first half of 2027, though JPMorgan has not announced a commercialization timeline.
The CCS read. JPMorgan is treating autonomous agents as infrastructure, not a finished product. The bank’s focus on policy engines and execution-time controls, rather than prompt engineering, signals that institutional payment systems will demand cryptographic or technical enforcement of agent boundaries, not trust. For blockchain-based treasuries and cross-border settlement, this framework may shape how custody providers and treasury platforms design API access and mandate delegation in the coming year.
Watch for JPMorgan’s disclosure of the multi-agent testing results in the first half of 2027. The key open question is whether agents operating under separate mandates will be permitted to retain shared memory between runs and, if so, what approval mechanisms will govern which inter-agent messages expand or modify an agent’s authority.
Original reporting: cryptopolitan.com


