D.C. Circuit upholds Pentagon's Anthropic supply-chain exclusion

A federal appeals court has upheld the Pentagon's exclusion of Anthropic from its supply chain in a 2-to-1 ruling, rejecting the company's challenge to procurement restrictions on its Claude AI models. The September 25, 2026 decision leaves in place an action covering the department's systems and contractors' work for it.

The D.C. Circuit's ruling turns on whether Anthropic's restrictions on military uses can count as a national-security supply-chain risk. The majority said they can under the law before it, even without a finding of malicious intent. A dissenting judge argued that the statute requires a narrower reading directed at hostile or deceptive interference.

The big change

• What changed: The Pentagon's use of supply-chain security powers to exclude Claude has survived review on the merits. The majority accepted that a supplier's openly declared safety restrictions can fall within this procurement law, even when the supplier acts with good intentions.
• Why it matters: The decision gives legal support to the Pentagon's position that it must be able to rely on AI performing the lawful tasks it requires. Anthropic's refusal to remove two use restrictions became grounds for exclusion from defense work covered by the action.
• What to watch: Anthropic says it is considering further review. Its separate California victory concerns a different statute, so the two rulings leave distinct legal questions for any further proceedings.

The contract dispute behind the exclusion

The dispute concerned the Pentagon's demand for permission to use Claude for all lawful purposes. Anthropic retained contractual prohibitions on lethal autonomous warfare and mass surveillance of Americans. The opinion recounts the company's position that current AI was insufficiently reliable for fully autonomous weapons and that mass domestic surveillance posed risks to democratic freedoms.

After negotiations failed, Defense Secretary Pete Hegseth issued a March 3 determination under the Federal Acquisition Supply Chain Security Act of 2018. The department ordered Claude removed from its systems as soon as practical, with 180 days as the outer limit, and prohibited contractors from using Anthropic products in their work for the department.

Friday's decision denied Anthropic's consolidated petitions for review. It follows an April decision refusing to pause the action while the case proceeded. The September opinion resolves the merits of this challenge; it does not establish how much of the ordered removal the Pentagon has completed.

Why the majority upheld the action

Judge Gregory Katsas wrote the opinion, joined by Judge Neomi Rao. They concluded that the department had adequate grounds to worry that restrictions embedded in Claude could prevent it from carrying out functions the Pentagon considered necessary and contractually permitted.

The majority cited earlier instances in which Claude refused government requests, including work involving classified materials and disease research. Anthropic said those problems involved safeguards in commercial models and had been resolved with the relevant government users. The judges treated the episodes as evidence that model training can enforce restrictions on use.

Anthropic told the court it cannot access, alter or shut down a model after delivery for use on the department's classified systems. The majority's reasoning did not require such remote control. It relied on restrictions already built into models and Anthropic's ability to build restrictions into later versions. The court also accepted the department's concerns about the limits of testing new models before deployment.

The legal question was whether that kind of control falls within the statute's definition of supply-chain risk. The majority read the definition to cover manipulation of a product's design or operation that denies or disrupts its use, without requiring a malicious motive. It acknowledged that Anthropic's intentions could be principled.

The judges also rejected Anthropic's constitutional claims. They found that the company received sufficient notice and an opportunity to challenge the exclusion after it took effect. On free speech, they recognized Anthropic's advocacy as protected but concluded that the exclusion resulted from its refusal to accept a contract term the department considered essential.

The dissent disputed the reach of the law

Judge Karen LeCraft Henderson focused on statutory language. She argued that words describing sabotage and malicious interference constrain the meaning of the broader phrase about manipulating a product.

On her reading, the law targets intentionally subversive or deceptive conduct. Anthropic's open enforcement of usage restrictions did not fit that category. She warned that the majority's interpretation allows the government to use a national-security designation against a contractor that honestly enforces restrictions the government dislikes, including restrictions previously agreed to.

The decision does not establish that fully autonomous weapons are safe, that a particular surveillance operation is lawful, or that Anthropic sabotaged a military system.

The California ruling remains a separate decision

In August, a federal district court in Northern California set aside a parallel designation under a different law, 10 U.S.C. § 3252. The D.C. Circuit reviewed the action under 41 U.S.C. § 4713.

The California statute refers to an adversary and to subversion; the law applied by the appeals court refers to any person and uses a broader definition of manipulation. The majority expressly accepted the California court's conclusion that Anthropic had acted without bad motive, while finding that such a motive was unnecessary under the statute it was applying.

For organizations assessing access to Claude, this ruling concerns the Pentagon's supply chain, including contractors' work for the department. It is not a ruling barring every federal AI contract or every commercial use of Anthropic products.

An Anthropic spokesperson told CNBC that the company disagreed and was considering further review. WIRED reported that the Pentagon had not provided detailed updates on its progress replacing Claude.