Appeals Court Upholds Pentagon's Anthropic Label; Dispute Is About AI Safeguards

Who decides how an AI model may be used: the company that built it, or the government that buys it? That question sits underneath the dispute between the Pentagon and Anthropic, the maker of the Claude models. It may matter to medicine, because hospitals face a similar question with every AI vendor they sign, though no source we checked ties this case to health-sector contracting.

A recent summary from our research desk, citing CNBC, reported that a federal appeals court upheld a Pentagon determination listing Anthropic as a supplier carrying supply-chain risk. It framed the label as the Pentagon applying ordinary procurement logic to software. We could not check that internal summary's framing externally. The ruling itself is confirmed by CNBC, the Washington Post, ABC and The Next Web.

The appeals ruling is confirmed. Judges Katsas and Rao formed the majority, and Judge Henderson dissented. The court denied Anthropic's petitions and rejected its claims that the designation was unauthorized, arbitrary, and a violation of free speech and due process. The majority said the department could reasonably find that Claude's built-in usage restrictions made it unreliable for military operations. So supply-chain law is the legal basis the court accepted, even though the factual trigger was a dispute over safeguards.

According to the public record, the dispute unfolded as follows:

• February 24, 2026. Defense Secretary Pete Hegseth threatened to remove Anthropic from the defense supply chain unless it allowed unrestricted military use of Claude. • February 26. Anthropic rejected the demand to drop its safeguards. Its stated red lines were mass domestic surveillance and fully autonomous weapons. • February 27. President Trump ordered federal agencies to stop using Anthropic models. • March 5. The Pentagon formally designated Anthropic a supply-chain risk. The company said it would challenge the move in court. • March 26. Judge Rita Lin of the Northern District of California issued a preliminary injunction, writing that the government's actions appeared to be classic illegal First Amendment retaliation. Amici supporting Anthropic at the trial court included Microsoft, the ACLU and retired military leaders. • August 27. Judge Lin granted Anthropic summary judgment on its First Amendment, due process and Administrative Procedure Act claims and vacated that designation. Anthropic lost on its claim that the presidential directive exceeded presidential power. The government said it would appeal.

• September 25. The D.C. Circuit upheld the separate designation made under a 2018 supply-chain security law. There were two designations under different statutes, challenged in two courts. The D.C. ruling did not reverse Judge Lin. One designation now stands and one is vacated. In dissent, Judge Henderson wrote that the law does not treat a contractor's honest and upfront enforcement of restrictions as a supply-chain risk.

Anthropic said it is considering its options, including asking the full D.C. Circuit to rehear the case. It says the designation has cost it billions.

This section is our inference, not a finding. Many AI tools come with usage terms. Some vendors forbid uses such as fully automated diagnosis without a clinician, or certain kinds of surveillance. Those terms are a form of safeguard.

This case tests whether a large buyer can pressure a vendor to drop such limits, and whether keeping them can be treated as a security risk. The answer may shape how much say AI companies keep over how their models are used once they are sold. Whether that reaches health systems is speculation on our part.

For patients, the practical point is simple. Safeguards on medical AI are only as durable as the contracts and laws that protect them.

If you use AI tools for health questions, or your doctor does, it is reasonable to ask what the vendor forbids and whether those limits are written into the contract. A safeguard that exists only as a company policy can change when a large customer pushes back.

• The single-label picture. There were two designations in two courts, and the rulings point in different directions. • The idea that the appeals ruling undoes the trial win. It does not; it concerns a different designation. • The weakest part of our argument. The link from this military dispute to hospital AI contracts rests on our inference alone.

This article was drafted with AI assistance that includes Anthropic's Claude models. We have kept to the public record and flagged claims we could not confirm.

• Whether the full D.C. Circuit rehears the case, and whether the government's appeal of Judge Lin's ruling proceeds. • Whether other federal agencies change their AI contracts in response. • Whether medical AI vendors keep or loosen their published usage limits under pressure.

Could Two Tidal Rhythms at Once Help Interdependent Microbes Survive? A Testable Idea About Timing, Not Just Energy Does the Pattern of Lost DNA Marks Decide Whether a Cell Can Recover? A Testable Idea, Rated 60 Percent