Judge rules Pentagon's Anthropic risk label unconstitutional

Asked to defend banning a company from the entire federal government, the Department of War produced four pages. Judge Rita F. Lin called the administrative record “slim” in her August 27 opinion, noting that the memorandum supplying “the entirety of the government’s rationale” post-dates two of the three actions it was meant to justify.

Washington’s original theory was sabotage: that Anthropic kept backdoor access to its models once they were running inside a national-security system. In litigation the government conceded that no such access exists, and that Claude is “no riskier to the national security than any other ‘black box’” model. That left one factor unique to Anthropic, which Defendants called trust, and they tied it explicitly to the company’s “increasingly hostile manner through the press.” Once the stated reason for a penalty is the target’s speech, the case stops being about national security and becomes a First Amendment case, which is how Lin decided it, alongside a Fifth Amendment finding that Anthropic never got the process it was owed before the penalty landed. The timing did not help the government either: days before the designation, Secretary Hegseth had proposed applying the Defense Production Act to Anthropic, which would have made it essential to national security rather than a threat to it.

Vendors selling AI into government should read the shape of this before the result. Anthropic lost its separation-of-powers claim outright, lost five agencies on its notice claim, and has no entered remedy yet, while a related designation remains undecided in the D.C. Circuit. What a supplier actually gains is procedural: an administrative record that has to survive being read by someone. That matters because this administration has already been intervening in AI litigation directly, building review gates around frontier models, and naming labs in public allegations. The paperwork is where a vendor’s defence now lives, and four pages was not enough of it.

Related

• The Justice Department told a federal judge that Minnesota’s nudification ban sweeps too broadly, without joining the case
• The government now gets a look at the most powerful AI models before the public does
• The White House says Moonshot AI trained its model on Anthropic’s, and accessed restricted Nvidia chips in Thailand