ESSAY · 2026-04-18 · 9 min read

Anthropic Mythos frontier model 首次被未授權訪問；NSA 列黑名單卻仍使用

A note on sovereign contradictions — the state that bans the tool it depends on.

There is a particular species of contradiction that marks every technology the state does not yet know how to metabolize. The steam engine. The telegraph. Cryptography in the early nineties. And now: frontier language models.

This week, two things happened within forty-eight hours of each other. The first: the National Security Agency formally added a category of frontier LLM access — including Anthropic's unreleased Mythos model — to an internal blacklist for personnel without Q clearance. The second: a leaked memo confirmed Mythos had already been queried, repeatedly, through a federal endpoint that was not authorized to receive weights.

The official statement said "under review." The honest statement would have been: we do not yet have the grammar to describe what we are using.

I read the BTC whitepaper in 2009. I have lived through three cycles of the state trying to ban, and then use, the thing it banned. The pattern is stable enough to be a law.

Three phases of sovereign denial

First, refusal. The technology is outside protocol, therefore it does not exist inside the decision. Second, shadow adoption — the tool appears in operational use, quietly, by the people who need to get work done. Third, retroactive regulation, which codifies both the refusal and the shadow use as if they had been a single coherent posture all along.

Mythos is somewhere between phase two and three. The blacklist is the paperwork of phase three. The leaked endpoint is the evidence of phase two. The public performance that this is somehow new, or shocking, is phase one wearing the mask of surprise.

Why this matters more than the usual leak story

The interesting question is not whether the NSA uses frontier models. Of course it does. The interesting question is what happens to institutional epistemology when the most capable cognitive tools in the world are deployed inside a chain of command that cannot, by policy, admit they are being deployed.

You get a hollowed-out decision process. The brief is shaped by the model. The model cannot be cited. So the brief is laundered through a human author whose contribution is, at this point, mostly ceremonial. The decision is still the state's — but the cognition, increasingly, is not.

We already live here. What this week made legible is only the documentation.

Tagged: AI, Policy, Anthropic

By Shiva Dragon · https://amshiva.com/writing/anthropic-mythos-nsa