Claude's coded restrictions support a Pentagon blacklist, appeals court rules

The court built its finding on a point Anthropic concedes. "As Anthropic admits, the company encodes restrictions into Claude that prevent the model from performing tasks that Anthropic wishes to prevent," the court wrote ^[6]. The panel held that the department had good reason to treat those coded limits as a supply-chain risk to itself and its contractors ^[3].

The board-deck version is that Anthropic lost and Claude leaves defense work. That version skips the question of posture. In Politico's report, the court found "ample support" for the department's conclusion ^[5]. ABC News described a court that declined to block the blacklisting ^[2]. A contractor planning around an interim refusal to block plans differently from one planning around a final ruling on the merits. Neither report says which this is, what the dissenting judge argued, or what contractors must remove and by when ^[8].

The contractor language in the opinion is tied to the department's information systems ^[5], and the blacklist as reported covers Pentagon contracts ^[1]. A firm running Claude inside a system it operates for the department falls within that language. A firm using Claude only in commercial work falls outside it. In my view the compliance risk is concrete for the first group and unproven for the second.

A skeptic would say this is one vendor's quarrel with one administration, and that contractors can wait for the litigation to finish. The court's stated reason cuts against waiting. It rests on restrictions a vendor codes into its own model ^[3], and any AI vendor selling a restricted model into defense work fits that description.

For a defense contractor with Claude in its stack, this quarter's decision is whether to start replacing it before the case ends. Starting now spends engineering time on a migration that a later ruling could make unnecessary. Waiting bets that the panel's 2-1 decision does not hold ^[1].

Next quarter's consequence follows from either choice. The replacement model will be judged on its usage policy as well as its capability. Under this court's reasoning, a vendor's coded limits count against it in a defense review ^[3].