Federal appeals court lets Pentagon keep Anthropic out of military contracts

The D.C. Circuit ruled that the Pentagon may exclude Anthropic from defense contracts because Claude's built-in safety constraints constitute a supply-chain risk under Section 4713. The decision hands the Trump administration a win in the fight over what counts as a "national security risk" and centers on the model's ability to refuse tasks the military asks it to perform, not on any remote control after deployment.

The court held that Claude's encoded safety rules could cause the model to reject tasks during a military operation abroad, when operational reliability is critical. The majority did not rely on the argument that Anthropic retains remote control after deployment, because the Pentagon clarified that the designation did not rest on that premise. Instead, the question was whether the mere existence of built-in constraints creates unacceptable uncertainty in the supply chain.

The risk, the court said, stems from the model potentially blocking legitimate military requests because of safety rules set in advance by the developer. The court accepted the Pentagon's position that the inability to predict or control such refusals in real time harms operational readiness. This is an expansive reading of "supply-chain risk" that goes beyond standard definitions of dependence on an outside vendor.

The panel acknowledged that Anthropic's AI safety work is protected speech, but ruled that the exclusion from tenders resulted from the company's rejection of specific contract terms, not from retaliation for that activity. The distinction leaves the door open for future claims if a direct link can be shown between the company's public positions and procurement decisions.

Unusually, the ruling coexists with a separate decision by Judge Rita Lin in California, which struck down the broader designation under Section 3252, including a government-wide ban and an attempt to bar subcontractors from working with Anthropic. The appeals court emphasized that Section 4713 defines "supply-chain risk" significantly more broadly than Section 3252, and explicitly noted it has no "dispute" with the California court's conclusion that Anthropic did not act with the improper motive required under the narrower statute.