Jamie Dimon told CNBC that JPMorgan Chase is building defensive tools against rogue agents, software that could act at scale and drain accounts faster than any fraud team could respond. The warning assumes a near future where autonomous programs hold real purchasing power inside a bank’s own systems. The harder question sits underneath it: if agentic commerce keeps coming regardless of what any one bank wants, can an institution defend against the very actors it is also racing to let transact?
Stripe answered with the opposite instinct. It is broadening its connection between AI shopping and checkout, so an agent completes a purchase the way any paying customer would, start to finish, no human in the loop. Laplace Intelligence works a different problem entirely. It red-teams robots and self-driving cars before they ever reach a public road, breaking the machine on purpose, in private, long before anyone trusts it with a sidewalk or a steering wheel.
Chase and Stripe hold opposite views of what an agent even is: a threat to contain, or a customer to serve. Laplace adds a third position neither bank nor payment company has fully claimed: test the thing to failure first, and only then let it move among people.
The split looks like a disagreement, but it is closer to a specification problem, and a training manual for payment engineers named it years before any of this made headlines. “The Anatomy of the Swipe,” by Ahmed Siddiqui and Nicholas Straight, walks through how 3D Secure 2.0 actually clears an online purchase: a push notification lands on the buyer’s banking app with roughly a five-minute window to approve it before the transaction expires.
The authors’ point is plain. A human cannot personally review every machine-initiated purchase, so the check has to live inside the transaction itself, timed to the moment money moves, never bolted on after the fact. Read against that, Chase’s wall and Stripe’s open rail stop looking like opposites. Both are attempts to answer the same five-minute problem, one by guarding the door and one by hosting the guest, and neither does any good without the other standing next to it.
That reframes the choice banks think they are making. A checkout rail with no check built into its own flow is an open line of credit wearing a costume. A defense raised against agents that never transact on it guards an empty room. Nobody here is trading safety for speed. Banking and payments are relearning, agent by agent, that the two only function when they are built into the same five-minute window, not weighed against each other in a boardroom.
The architecture that wins this will not belong to whoever raises the strongest wall or opens the fastest checkout. It will belong to whoever makes the approval notification, the red-team report that grounded a car before its first public mile, and the rail that settles the bill all answer one question at the instant the money moves: who checks this next? An agent is not dangerous because it can move fast. It is dangerous only in the stretch of road where nothing is built to ask it what it is doing.