An administration letter to Anthropic locked the NSA out of Mythos too

A letter that cut off its own intended customer

The reporting centers on Mythos, an Anthropic model framed in near-apocalyptic terms for offensive security. One line quoted from the article in the thread sets the tone: Mythos “broke into almost all of our classified systems, not in weeks, but in hours.” The White House and intelligence officials had pushed a classified contract that would let the N.S.A. use the technology for intelligence analysis and detecting new computer vulnerabilities. That contract was never finalized, and some Pentagon officials wanted the agency to find a way to work with other models instead.

The twist: the government issued a directive to Anthropic that, by its breadth, pulled access from everyone, down to Anthropic's own employees. The agency that most wanted Mythos ended up locked out alongside the public.

The thread reads it as an own-goal, not a seizure

The top objection was that the state could simply compel delivery. The correction came fast, on the legal lever and on the politics of using it.

“The DPA only gives that power to the President…” — a growth-equity VC citing the Defense Production Act

That reframes the whole episode as procedural. One commenter put the absurdity bluntly.

“NSA didn't have to if the administration had thought about it for 30 seconds before sending their letter. It's a stupid own-goal.” — a longtime user on the directive's blast radius

Others attacked the premise that losing weights matters. The argument: the model alone isn't the weapon.

“the success of mythos isn't from model weights, it's from the harness and toolset it has access to” — a commenter on what makes the tool dangerous

That thread continued into how an audit tool actually breaks things, against skeptics pointing to the modest confirmed curl finding as proof of hype.

“Mythos's great strength was finding multiple vulnerabilities and chaining them together to break a whole system.” — a former ML founder, ex-CTO on AI for policy

Underneath sat the contradiction both parties live with: the government calls the model a supply-chain risk while threatening to compel it, and the vendor calls it nuclear-grade while shipping it commercially.

“The model is both a supply chain risk and also we'll DPA you if you don't give it to us.” — a commenter mapping the horseshoe

The practical residue isn't who has the model. It's that a usage policy written to contain a capability can revoke it faster, and more indiscriminately, than any adversary could.