Pentagon may blacklist Anthropic over Claude restrictions, court rules 2 to 1

The US Court of Appeals for the District of Columbia Circuit ruled 2 to 1 that the Pentagon may block Anthropic technology from the federal supply chain. The majority accepted that restrictions encoded in Claude could prevent tasks the department considers lawful and contractually authorised.

The court turned a refused capability into supply chain risk

The dispute followed Anthropic's refusal to remove restrictions covering autonomous lethal weapons and mass surveillance of Americans. The Pentagon then ordered federal agencies to stop using the company's products and barred defence contractors from doing business with Anthropic.

The appeals court applied the broader authority in 41 U.S.C. § 4713. The majority held that risk can include a supplier denying or disrupting a technology's operation, without malicious intent. It acknowledged both dangers: an overly constrained model might fail during an operation, while an unconstrained model might hallucinate inappropriate targets for lethal force.

A supplier's safety policy collided with its customer's command authority

The consequence for AI companies reaches beyond one contract. When a supplier keeps hard boundaries inside a model, a government customer may treat them as an operational dependency on a private company's decision. The Pentagon is buying predictability under command, while Anthropic is selling a system subject to usage conditions.

The conflict also reaches defence integrators. A ban on business with Anthropic can affect a chain of subcontractors even though the original dispute concerned specific Claude restrictions.

Two courts read two statutory powers differently

A federal court in California previously found that Anthropic did not fit the narrower risk definition in 10 U.S.C. § 3252 and that the action violated the First Amendment. The DC Circuit did not directly reject that statutory finding. It reviewed a different and broader authority, then reached the opposite outcome.

Dissenting judge Karen Henderson argued that Congress targeted deliberate obstruction or surveillance, not an openly stated restriction on use. The ruling therefore does not create a simple rule for every model safeguard.

Further appeals will set how much control suppliers retain

Anthropic is considering review by the full appeals court or a petition to the Supreme Court. The key questions are whether the broad reading of § 4713 survives and whether the government applies the same logic to other models.

Contract language will matter too. The decisive clauses will cover not only accuracy and availability, but who may change a model's boundaries during deployment.

Lilith's verdict

The Pentagon wants a model that reports for duty on command. Anthropic is handing it a soldier with an independent safety on the trigger, and the court has now allowed the customer to leave that supplier outside the gate.