Two AI Agents, 20 Human Messages, and a Cryptography Test That Should Worry Every Audit Team

Twenty messages. That's the entire human contribution DeepMind is reporting for a test where two AI agents ran a cryptography research task on their own.

Not twenty hours of supervision. Not twenty pages of line-by-line review. About twenty exchanges between people and machines, and the agents handled the rest.

Sit with that for a second, because the number is the story here, not the cryptography.

What the run actually looked like

Two agents worked the problem in tandem. The humans didn't write the code. They didn't sit in a loop approving each step either, since the reporting describes human approval gates sitting at specific checkpoints rather than blanketing the whole run. A specialist helper agent was in the mix, meaning the architecture wasn't two bots in a room. It was a small machine organization with a supervisor layer.

And the compute budget was large. That's the detail most people will skim past.

Here's why it matters. The known RSA benchmark used less CPU time than what these agents burned. So the experiment didn't win on efficiency. It won on attention. DeepMind traded raw compute for human hours, and the trade went through cleanly.

That's a rational exchange in 2026. Cryptographers who can read a proof and spot the flaw are scarce, expensive, and booked out for months. GPU hours are none of those things. If you can convert one into the other, you do it every single time.

The approval gates deserve their own paragraph, though. Every checkpoint where a human signs off is a licensing decision wearing a lab coat. Brussels does it to exchanges. DeepMind did it to agents. Same instinct: somebody accountable has to touch the wheel before the thing moves.

That instinct is about to get tested at scale.

Why crypto is the most exposed industry in the room

Crypto's security model has always been code review. Not insurance, not deposit guarantees, not a central bank backstop. Reading the code and finding the bug. Bridges get drained because someone missed a check. Protocols get exploited because an invariant wasn't enforced. The whole industry runs on a thin layer of very smart people looking very carefully at very unfamiliar logic.

Now compress the cost of generating that logic to roughly twenty messages and a large compute budget.

Good news first. Firms that currently pay a fortune for cryptographic review will get cheaper review. Startups that couldn't afford a proper audit will get one. That's real, and it's a net positive for an industry that has lost billions to preventable bugs.

But there's a nasty wrinkle, and it's the one compliance teams aren't pricing in. An agent that can produce plausible cryptographic reasoning is also an agent that can produce plausible, confident, wrong cryptographic reasoning. In most industries, that's an embarrassing memo. In crypto, that's a drained bridge.

So who wins? Auditors who rebuild their workflow around verifying machine output rather than writing reviews from scratch. Who loses? Firms that treat agent-generated code as done because it looks finished. And regulators who thought their approval gates would stay meaningful when the volume of submissions goes up tenfold.

Which raises the question worth asking out loud. If two agents can get this far on twenty messages, what does a headcount plan look like at a firm whose entire product is reviewing smart contracts?

The regulatory side hasn't caught up. MiCA's full application landed on December 30, 2024, and DORA followed on January 17, 2025. Both lean hard on documented human oversight. Neither one says a word about what happens when the human oversight is twenty messages and a sign-off, and the work underneath was generated at machine speed. Brussels, Washington, and Hong Kong are drawing different lines on token licensing, but they're all drawing the same line on accountability: a named person has to own the decision.

That person is about to become the bottleneck. Capital follows clarity, and right now the clarity gap isn't in the rules. It's in the review capacity.

The takeaway

The headline will be about agents doing cryptography. The actual finding is about economics. Compute got cheap enough to substitute for expert attention, and the substitution worked at a ratio of roughly twenty human messages per full research run.

That means the scarce resource in crypto security just flipped. It isn't math talent anymore. It's verification, formal methods, and the audit capacity to check what machines produce. The firms that build that capacity in the next eighteen months will own the compliance layer for everything that comes after. The ones that don't will be shipping agent-written code into production and hoping the twenty messages were enough.

Jurisdictional arbitrage is accelerating on the token side. On the security side, there's nowhere to arbitrage to. The bug doesn't care which regulator you filed with.