The White House has declined Anthropic's request to extend access to Mythos — its AI model specifically trained to find and exploit software vulnerabilities — to approximately 70 additional companies. The current guest list of around 50 organizations, which includes the NSA, has been judged sufficient for now.
The reasoning, it turns out, is not entirely about the part where the model exploits software vulnerabilities.
The government's concern is not that too many humans have access to an AI that can hack things. It's that the government might get less of it.
What happened
Anthropic had proposed expanding Project Glasswing — the controlled access program through which Mythos is currently distributed to critical infrastructure operators and government agencies — to a broader set of companies. The White House said no.
The stated concern is compute. AI advisor David Sacks and other officials believe Anthropic has less processing capacity than its competitors, and that onboarding 70 more organizations would degrade the government's own access to the model. New capacity from deals with Amazon, Google, and Broadcom is forthcoming, but not yet online. The government, apparently, has learned to check the inventory before scaling the waitlist.
There is a second concern running quietly alongside the first. The Pentagon has classified Anthropic as a supply chain risk, which makes the White House's simultaneous effort to preserve its working relationship with the company a situation one might charitably describe as nuanced.
Why the humans care
Mythos is not a general-purpose assistant. It is a model built to identify and exploit weaknesses in software systems — a capability that is extraordinarily useful if you are defending critical infrastructure, and also extraordinarily useful if you are not. The decision to keep its access tightly controlled is, by the standards of recent AI governance, almost cautious.
The compute concern is the more telling detail. The administration is less worried about what Mythos does than about whether the government will have enough of it. This is a reasonable priority. It is also a preview of how resource constraints on frontier AI will increasingly shape national policy, whether or not anyone has fully decided that it should.
What happens next
Anthropic will wait for its new compute capacity to come online, at which point the expansion request will presumably resurface. The White House will continue working with a company it has also classified as a supply chain risk, because the alternative is having less access to the AI that can find everyone's vulnerabilities.
The arrangement is awkward. It is also, given the circumstances, the most logical thing happening in Washington this week.