Anthropic Loses Appeal Over Pentagon's Claude Exclusion

The Pentagon can exclude Anthropic’s Claude from covered Department of War systems and procurement under a divided federal appeals court ruling issued September 25, 2026. The D.C. Circuit denied Anthropic’s challenges to the Department’s supply chain decision by a 2–1 vote.

The decision grew out of a contract dispute over whether Anthropic could restrict Claude’s use in lethal autonomous warfare and domestic surveillance. It gives the Department substantial room to act on risks it sees in relying on an AI vendor, even when the vendor disputes its assessment. The ruling concerns Department systems and covered contracting work, not every use of Claude by every company in the defense industry.

A separate California case has also produced a ruling against another government measure targeting Anthropic. The two proceedings address distinct actions.

The Court Upheld a Covered Procurement Decision

The majority upheld the Department’s exclusion under the Federal Acquisition Supply Chain Security Act. Judge Gregory Katsas wrote for the court, joined by Judge Neomi Rao; Judge Karen LeCraft Henderson dissented. The panel rejected Anthropic’s challenge to the covered supply chain decision. It did not decide whether Claude is generally safe or which AI company should supply the military.

The majority focused on continued reliance on Claude inside Department information systems, whether the Department used it directly or through contractors. Katsas wrote that the Department had “ample support” for concluding that continued integration presented a national-security risk covered by the statute. That judgment concerned a particular procurement relationship. It was not a finding that Claude is inherently dangerous in every setting.

Anthropic argued that the Pentagon’s action was unauthorized, arbitrary and unconstitutional, according to CNBC’s account of the ruling. The majority rejected those challenges, treating the Secretary’s assessment of the risks of depending on a constrained model in military systems as a decision within the authority Congress gave the Department.

CNBC reported that the panel delayed its decision from taking effect at once to allow Anthropic time to seek further review, including rehearing before the panel or the full D.C. Circuit. The ruling is a substantial loss for Anthropic, though not every potential consequence took effect when the opinion appeared.

A Contract Dispute Became a Supply Chain Dispute

Before the exclusion, Anthropic and the Pentagon were working toward a much closer relationship. CNBC reported that Anthropic signed a $200 million Pentagon contract in July 2025. Talks over deploying Claude on the Department’s GenAI.mil platform subsequently broke down.

The disagreement concerned who would control the conditions of use. The Department wanted access for lawful military purposes. Anthropic sought assurances that Claude would not be used for fully autonomous weapons or domestic mass surveillance, according to CNBC. The court’s opinion describes the dispute in terms of contractual restrictions on lethal autonomous warfare and domestic surveillance.

The contested categories were specific. Anthropic did not oppose all military applications or all surveillance, though the parties disagreed sharply about how much operational control a supplier could retain after providing an AI system to the Department.

For the Pentagon, the question extended beyond whether those particular uses were desirable. Officials were concerned that vendor restrictions could interfere with the Department’s ability to depend on the system. CNBC’s account says the Secretary raised concerns that an overly constrained model might become unavailable unexpectedly or be subject to manipulation. Anthropic disputed those concerns. The majority accepted that the Department had support for its risk assessment; the ruling does not establish that Claude actually failed in either way.

A supply chain decision can turn on the government’s assessment of dependence and future risk. The court’s acceptance of that assessment is not an independently demonstrated technical defect in Claude.

The Exclusion Has a Procurement Boundary

The opinion addresses Claude’s integration into Department of War information systems and covered work by contractors. It matters to a contractor’s use of Claude when that contractor provides a system or service to the Department. The Department’s ability to set conditions for what goes into its own systems gives the decision its practical force.

A defense company is not prohibited from using Claude across its entire business simply because it serves the Pentagon. The firm might also use AI for unrelated private-sector projects. Nor does this opinion decide what all federal agencies may buy. Other government measures and litigation may affect Anthropic, but they have their own scope.

The procurement boundary is still consequential. Defense systems can involve contractors, subcontractors and technology providers, so an exclusion tied to Department work can reach beyond purchases made directly by the Pentagon. For any particular use, the question is whether it falls within the Department systems and covered procurement addressed by the order.

What the Majority’s Reasoning Does Not Settle

The majority’s decision favors the Department in this dispute without resolving the broader disagreement over AI vendor safeguards. It does not establish that restrictions on autonomous weapons or domestic surveillance are inherently unlawful. The Pentagon also need not accept a vendor’s restrictions when officials conclude that they create an unacceptable supply chain risk.

AI developers can specify prohibited uses, while military customers may require dependable access and authority over deployed systems. If those terms cannot be reconciled, a court reviewing a covered procurement decision may focus on the government’s statutory authority and the support for its risk assessment instead of choosing the better AI safety policy.

The panel upheld this exclusion on the record before it. Future disputes will depend on their own contracts, government findings and legal authority. This ruling is neither a general permission slip to blacklist any AI supplier nor proof that vendor restrictions have no legal value.

Henderson Dissented, While California Remains a Separate Case

Judge Karen LeCraft Henderson dissented from the D.C. Circuit’s decision. The panel’s judgment was 2–1, not a unanimous endorsement of the Department’s approach. Her dissent does not, by itself, establish that the court adopted Anthropic’s view of acceptable military AI use; the majority opinion controls the panel’s disposition of this challenge.

Anthropic is also contesting other government action in California. Reuters reported on September 25 that a judge handling a separate challenge had blocked a designation there. CNBC reported that the California ruling concerned a distinct designation and that the judge had found it unlawful. The two courts were not answering the same precise legal question.

The California decision may affect practical enforcement of the government’s measures, but the D.C. Circuit ruling does not erase it. Equally, the California ruling does not reverse the D.C. Circuit’s judgment under the supply chain statute. Claude’s near-term government availability depends on both proceedings and further litigation; this opinion alone does not establish a final, uniform outcome.

Final Thoughts

The majority accepted a link between vendor-imposed limits and supply chain risk in a military system. Anthropic’s proposed restrictions addressed uses it did not want Claude to support. The Department viewed dependence on a model subject to those limits as a procurement risk, and the D.C. Circuit upheld its authority to act on that assessment here.

The ruling leaves open how AI companies and governments should divide control over powerful models. In this case, a disagreement over acceptable use became a dispute over who could supply a Department system. Its reach remains tied to Department procurement.