Verification Is the New Frontier

This weekend Dario Amodei published “We Must Pace the Frontier.” The other major labs signaled agreement within hours, which tells you how long this has been coming. Whatever framework gets formalized, the people building the most capable AI systems in the world just told the market that they can no longer verify their own systems fast enough to keep racing.

‍

We read this as a finance story.

‍

The labs didn’t discover a new problem. Past a certain level of stakes, trust cannot be asserted — it has to be proven, continuously, by someone other than the party making the claim. Finance has organized itself around that principle for a hundred years. It’s why clearinghouses exist. The labs, in their own way, just reinvented the audit function.

‍

Look at the first concrete commitment on the table. It isn’t a compute cap. It’s embedded third-party evaluators with employee-level access, checking that labs actually do what they say. Humans, checking — and for model development that’s probably the right mechanism, and it will work for longer than infrastructure people like us want to admit. Where it breaks is the next phase: agents acting at machine speed against real systems, and increasingly against real money. Treasury, payments, settlement, custody. No bank will let an autonomous agent move value because an evaluator reviewed the lab’s safety culture last quarter. They will want proof per action, and they will want to hold that proof themselves.

‍

Proof per action is infrastructure, and it’s the infrastructure we build. An agent action running through Blockdaemon’s control plane gets evaluated against the institution’s own policy, then authorized (or kicked up to a human), enforced through MPC threshold signing, and recorded as a signed attestation the institution can export: which agent, which policy, what time. The institution holds the artifact. If a regulator asks, they don’t have to take a vendor’s word for what happened.

‍

The control plane runs on the same rails that support 400+ institutional customers and over $110B in digital assets — six years, zero key compromise events. DTCC, NYSE, and LSEG don’t work with us because we move fast. Speed was never the pitch. Everything we do is verifiable, and that property is suddenly the one the entire AI industry is trying to establish.

‍

So our position on the slowdown: it isn’t a retreat, it’s a repricing. For a decade the AI race rewarded velocity. In regulated finance, the binding constraint from here is governance — the models are already good enough for most of the workflows that matter. What’s been missing is the layer that lets a risk committee, an auditor, and eventually an examiner say yes.

‍

The labs are pacing the frontier so safety can catch up. Our job is to make sure that when institutions are ready to deploy, the trust layer is already there.

‍

It is.

‍

Blockdaemon’s agentic control plane brings policy enforcement, MPC-based key control, and cryptographic attestation to AI workflows. Talk to our team.