The National Security Agency is quietly running Anthropic's Mythos Preview, the most capable model the company has ever built, two sources told Axios. That would be a routine procurement story, except for one inconvenient detail: the NSA sits inside the Department of Defense. And the Department of Defense has spent the last three months telling a federal court that Anthropic is a supply chain risk the government should not do business with.

Pick one. You cannot hold both positions at the same time, which is what makes this week's revelations so telling about how the federal government actually makes decisions about frontier AI.

What Mythos Actually Is

Mythos is not a chatbot upgrade. The UK AI Security Institute evaluated it as "substantially more capable at cyber offence than any model previously assessed." JPMorgan Chase chief executive Jamie Dimon told analysts on his earnings call this week that the model has already uncovered thousands of vulnerabilities in corporate software the bank tested it against. Dimon, not prone to tech hype, conceded that AI is now "making cybersecurity worse" in the short term.

Anthropic is releasing it to a small, curated group under a program called Project Glasswing. JPMorgan is in. A handful of federal agencies are in. Most of the world is not.

The Contradiction

In February, President Trump posted that his administration "will not do business" with Anthropic, and the Pentagon followed up by designating the company a supply chain risk. That designation is now the subject of an active lawsuit from Anthropic, which argues the label was designed to cripple its federal revenue.

And yet, per a White House OMB memo Bloomberg obtained this week, Treasury, Homeland Security and the Department of Defense itself are lined up to use a version of Mythos. The NSA already is. On Friday, White House chief of staff Susie Wiles and Treasury Secretary Scott Bessent met with Anthropic chief executive Dario Amodei. The White House called it "productive and constructive." That is the language governments use when they have realized they need the thing they publicly called dangerous.

Why This Matters Beyond One Company

The story here is not whether Anthropic gets its contracts back. It almost certainly will. The story is that US AI policy is now being written inside closed-door meetings between frontier labs and the intelligence community, then backfilled with public statements that contradict the private operational reality.

That is a governance model. It is not a good one. The public sees "supply chain risk." The NSA sees a tool that can find zero-days faster than any red team on the planet. Both sentences are true. Only one of them drives procurement.

Yoshua Bengio, one of the three so-called godfathers of AI, flagged the deeper issue this week. The real problem, he argues, is not whether Mythos is deployed. It is that a single private company now gets to decide which governments and which corporations get access to a tool that materially changes the offensive cyber balance. Anthropic is picking winners. Everyone else, friendly allied nations included, is told to wait.

This is what AI concentration looks like in practice. Not a single superintelligence running the world. A handful of labs deciding, case by case, who gets the sharpest knife.

First reported by Axios, with additional reporting from Bloomberg, BBC, and PBS NewsHour.