Defense sellers should plan Pentagon systems without Claude

Picture the engineer whose document-triage tool for a defense program calls Claude. On Monday she has two court rulings in front of her that point different ways, and a program manager who wants one answer.

Both rulings can stand at once because Hegseth designated Anthropic under two different laws ^[14]. The one the DC Circuit upheld ^[10] rests on the Supply Chain Security Act. "In doing so here, the Secretary did not transgress any limits on his authority under the Supply Chain Security Act or the Constitution," the court wrote ^[13]. Judge Lin's ruling on the other law also blocked the government-wide ban on Anthropic and Hegseth's order barring military contractors from doing any business with the company ^[6].

The majority's reasoning is about how the product behaves. "The Secretary raises the deeply sobering prospect of overly constrained AI models shutting down unexpectedly and thus causing important military operations to fail," the majority wrote in an opinion by Judge Gregory Katsas ^[11]. The ruling set Anthropic's concern beside it: "Anthropic raises the deeply sobering prospect of unconstrained AI models hallucinating inappropriate targets for lethal military force" ^[12]. It then left the choice to the executive branch, saying Trump and Hegseth "must determine how best to balance the competing risks" ^[15].

The feature Anthropic pitches is the one the Pentagon objected to. Anthropic presents its refusals as a safety position, saying AI is not yet reliable enough to be used safely in autonomous weapons ^[19]. Hegseth argued the same restrictions could compromise military operations ^[7]. According to Ars Technica, the judges held that the government could blacklist Anthropic for withholding the features even if the company had no malicious intent ^[3].

Teams tell themselves the San Francisco ruling means this unwinds. The record on the supply-chain designation runs the other way. It dates to March ^[20]. The same appeals court refused Anthropic an emergency stay in April ^[16] and upheld the designation on Friday ^[1]. An Anthropic spokesperson told CNBC, "Another federal court has already held the government's parallel designation unlawful" ^[9]. The statement is accurate, and it refers to the other law ^[14]. Commerce Secretary Howard Lutnick recently said the administration and Anthropic are "in tune" ^[18].

The panel found "ample support" for the Pentagon's view that integrating Anthropic's technology into military systems could pose a national security threat ^[4]. The reports do not spell out how far the upheld designation reaches into a contractor's internal tools.

I'd plan for Claude to stay out of anything delivered into a Pentagon system for as long as the designation stands. The tradeoff is paying now to re-test prompts and evaluations on a second model, against a designation Anthropic is still contesting.

Sort each program on two axes: whether Claude sits inside what the military receives or only in internal tooling, and whether swapping the model takes a sprint or a quarter of re-testing. Inside the deliverable and slow to swap is where the upheld finding lands ^[4], and the migration starts now. Inside and quick to swap, a tested second model named in the proposal covers the risk. Internal and slow to swap, the protection is Judge Lin's block on the contractor-wide ban ^[6]. Internal and quick to swap is the one quadrant where waiting for further review costs little.