Washington appeals court upholds Anthropic’s exclusion from Pentagon contracts

The federal appeals court for the District of Columbia, the D.C. Circuit, rejected Anthropic’s challenges to its exclusion from the Pentagon’s supply chain on September 25, 2026. Issued by a two-to-one vote, this ruling on the merits upholds the measures taken against the provider of Claude and its products under Section 4713 of Title 41 of the U.S. Code.

The exclusion applies to military contracts and systems covered by these measures. Civilian uses of Claude fall outside the scope of the ruling, which orders no changes to the model.

The Washington decision concerns proceedings separate from the California ruling in Anthropic’s favor in August. The two courts examined measures taken under different legal authorities.

Two proceedings under different legal authorities

The proceedings reviewed in Washington stem from the administrative decision of March 3, 2026, of which Anthropic was notified on March 4. The D.C. Circuit had declined on April 8 to temporarily stay the measures before reviewing them on the merits. The September 25 ruling now resolves the challenges themselves.

On August 27, Rita F. Lin, a judge in the U.S. District Court for the Northern District of California, had struck down separate measures targeting Anthropic, including a designation under Section 3252 of Title 10 of the U.S. Code. She had also ruled on constitutional rights.

Her ruling had blocked the broad government ban and the order preventing military partners from doing business with Anthropic, including outside military contracts.

The California ruling remains in effect, ABC News reports. The Washington ruling does not overturn it. The difference lies in the legal authorities and measures examined in each proceeding.

Contractual restrictions and technical refusals

The mechanism used by the Pentagon comes from the Federal Acquisition Supply Chain Security Act of 2018, a federal law on supply chain security in public procurement. Under certain conditions involving national security risks, Section 4713 authorizes the exclusion of a supplier from contracts and subcontracts related to work for the agency. Judicial review is governed by Section 1327 of Title 41.

This supply chain includes the tools, suppliers and service providers on which military systems depend. The exclusion can therefore apply to the use of Claude in a service provided to the Pentagon by another company.

Anthropic was already working with the U.S. defense establishment before the dispute. In his February 26 statement, Dario Amodei cited intelligence analysis and operational planning among Claude’s uses.

The company maintains two contractual restrictions concerning fully autonomous lethal warfare and mass surveillance of Americans. The first targets systems that select and engage targets without human involvement in that decision. Anthropic considers current models too unreliable for this use.

The Pentagon is demanding a clause allowing all lawful uses and takes the view that a private supplier should not have a veto over its operations. It presents this clause as a way to reduce the risk of refusal during an operation. A model’s acceptance of a request, however, proves neither its reliability nor the lawfulness of the use.

Contractual clauses define permitted uses. Safeguards built into the model affect its responses and refusals, while versions adapted for the government correspond to specific deployments. A contractual restriction and an automated refusal therefore impose different constraints on a military integrator.

The majority finds an operational risk

The opinion by Gregory G. Katsas, joined by Neomi Rao, holds that the Pentagon could find a risk of refusal or disruption of military tasks. The majority relies on documented refusals involving versions of the model and on contractual uncertainty. This assessment concerns an operational risk, without establishing military sabotage.

In the majority’s view, the law can cover effects involving manipulation or disruption even without malicious intent. The judges do not characterize Anthropic as malicious and do not question its intentions to protect privacy and artificial intelligence safety.

The reconsidered administrative decision does not rely on any control Anthropic might exercise over classified systems after delivery.

Karen LeCraft Henderson disputes this interpretation in her dissent. In her view, restrictions honestly disclosed by a supplier do not correspond to the risk of sabotage or subversive conduct contemplated by Congress. She also fears pressure on other suppliers’ safety safeguards, an effect that remains hypothetical.

The D.C. Circuit also rejects Anthropic’s constitutional arguments concerning freedom of speech and procedural safeguards.

A scope of application clarified as early as March

The Pentagon had already set out how the measures would apply in a March 6 memorandum and clarified their scope on March 19, as the September 25 ruling recalls.

For subcontractors using Claude in a military service covered by the exclusion, switching suppliers could entail new validations and replacement costs. These consequences depend on the contracts and systems concerned; no amount has been established.

Anthropic says it is reviewing its options, including rehearing by the full court, Reuters reports in a dispatch published by CNA.