Appeals Court Keeps Pentagon's Supply-Chain Risk Label on Anthropic—for Now

A federal appeals court in Washington, DC on Wednesday denied Anthropic’s emergency motion to lift the Pentagon’s “supply-chain risk” designation, letting the label remain in force while litigation continues. The three-judge panel ruled that Anthropic “has not satisfied the stringent requirements” for an immediate stay. That decision clashes with a March 26 order by U.S. District Judge Rita Lin in San Francisco, who called the move “Orwellian” and issued a nationwide injunction barring the government from cutting off Anthropic’s federal contracts. The dueling opinions leave the AI company in legal limbo: the label exists, but—at least for the moment—it cannot be enforced against the firm or its partners.

The Department of Defense notified Anthropic on March 3 that it was formally branded a supply-chain risk—the first time such a tag has been applied to a U.S. company, according to BBC and Politico. Officials cited the company’s refusal to grant unfettered military use of its Claude model, particularly Anthropic’s explicit ban on deploying Claude for mass surveillance or autonomous weapons. Defense Secretary Pete Hegseth escalated the dispute on social media, warning contractors to sever ties “effective immediately.” The Pentagon’s March letters also instructed prime contractors to scrub Anthropic software from their systems within 90 days, a step Mayer Brown notes could cascade through every tier of the defense supply chain.

Judge Lin found the government’s action likely violated Anthropic’s Fifth Amendment right to due process and First Amendment right to speak through its usage policies. In a 28-page ruling, she wrote that “nothing supports the Orwellian notion that disagreeing with the government turns a company into a national-security threat.” Courthouse News reports that the judge emphasized the lack of individualized findings against Anthropic and the absence of any procedure for the company to contest the label. Because the designation was issued without notice or hearing, Lin concluded the firm faced irreparable harm from the immediate freeze on federal work.

The appeals panel did not reach the constitutional merits. Instead, it applied the traditional four-factor test for emergency stays and decided Anthropic failed to show it would suffer irreparable injury in the short term. Bloomberg notes the judges highlighted that the lower-court injunction already protects the company from immediate contract terminations, so the risk label itself does not presently inflict concrete harm. Wired adds the court gave the government seven additional days to appeal the California injunction, setting up a potential circuit split that could land the case at the Supreme Court.

With the label hanging over it, Anthropic’s cloud partners—Amazon Web Services and Google Cloud—must decide whether to treat Claude as restricted software. FedScoop reports that prime contractors such as Lockheed Martin and Raytheon have already begun auditing codebases for Anthropic models, even though Judge Lin’s order shields them from immediate termination. Smaller AI vendors now see a chilling effect: fearing similar retaliation, several startups have quietly removed usage-policy language that bars military applications, according to industry attorneys cited by Mayer Brown. Venture investors are also adding “Pentagon risk” clauses to term sheets, potentially steering capital away from safety-focused labs.

The Justice Department has until mid-April to appeal the California injunction; if it does, the Ninth Circuit will weigh whether to keep Judge Lin’s order in place. Meanwhile, Anthropic’s broader constitutional challenge is scheduled for a district-court bench trial this summer. Legal scholars interviewed by NPR expect the case to clarify how much control the government has over private AI models that merely transit federal systems. A definitive ruling could set precedent for other frontier labs like OpenAI and Anthropic rival xAI, whose models also operate under usage policies that limit defense applications.

At its core, the dispute is about who sets the rules for military use of general-purpose AI. The Pentagon argues that any restriction on its ability to deploy commercial models creates a strategic vulnerability vis-à-vis China. Anthropic counters that unchecked military deployment could accelerate arms races and erode democratic oversight. The standoff echoes Cold War-era battles over cryptography export controls, but with far larger commercial implications because today’s models power everything from logistics software to battlefield chatbots. How the courts resolve this tension will likely influence pending congressional efforts to create a federal AI licensing regime.