Did the Pentagon Prove Anthropic Was a Supply-Chain Risk?

What the government could not show

Anthropic's position in the stalled contract talks was blunt. Frontier models were not ready, the company said, for mass domestic surveillance or for lethal targeting and firing decisions. The Pentagon's reply was that a private vendor does not get to dictate lawful military use.

That contract fight became something larger. Officials cast Anthropic as untrustworthy: opaque models, privileged access for the builder, guardrails that might shift after delivery, and executives willing to question how Claude was being used in an ongoing operation. DOJ counsel James Harlow told Lin that frontier AI is not a rifle you can strip and inspect. Each update, he argued, is another chance for an undisclosed moral rule to land inside a warfighting system.

Lin pressed for articulable facts. She said she saw no evidence Anthropic could alter a delivered model or flip some kind of kill switch. The kill-switch theory, central to the supply-chain story Washington wanted to tell, arrived without the proof a court can weigh.

Speech as evidence of enmity

The sharper constitutional problem was how the government treated Anthropic's public criticism. Lin called the idea that criticizing the Department could justify labeling a contractor a supply-chain risk really troubling. On the hearing record, that logic starts to look like a warning to every vendor watching: argue with us in public, and we may treat you as an adversary under procurement law.

Anthropic's counsel Michael Mongan argued the challenged acts were exercises of sovereign power, not ordinary day-to-day contract management, and that the Pickering balancing test for employee speech does not stretch this far. Even if it did, he said, the shifting rationales, from post-deployment sabotage fears to pre-deployment control once Anthropic showed it lacked a remote kill switch, look like pretext.

The timing matters for the frame. The risk-analysis memo DOJ leaned on was assembled after presidential and Pentagon directives in late February. Public dispute, then paperwork. Lin's skepticism tracked that order.

Meanwhile the Department told the court it is winding Anthropic products off defense networks by September 30, even as Lin's injunction still blocks the formal ban. Agencies may choose other vendors. What they should not get is a doctrine that converts a safety refusal, and press about it, into proof that the lab is an enemy in the supply chain.

Lin took the motions under submission. The answer to the title, on the record she described from the bench, is no. A government that wants unrestricted use of a frontier model can walk away from the deal. Inventing a kill switch it cannot document, then punishing the vendor for saying so out loud, is a different project. Courts exist to notice the difference.