This summary is based on the United States Court of Appeals for the District of Columbia Circuit’s published decision in Anthropic PBC v. United States Department of War, decided on September 25, 2026.

The case concerns the decision by the Department of War to exclude Claude, an artificial-intelligence product developed by Anthropic PBC, from its supply chain under the Federal Acquisition Supply Chain Security Act of 2018. The Department made this determination after Anthropic refused to relax contractual prohibitions preventing Claude’s use in lethal autonomous warfare or domestic surveillance, which were contract terms deemed essential by the Department.

The court’s opinion, authored by Circuit Judge Katsas, rejected Anthropic’s challenges to the Department’s actions. The company had argued that the exclusion was arbitrary, unauthorized by the statute, and unconstitutional (including due process and First Amendment claims). The court found that the Department’s decision had sound statutory and factual support, and that the constitutional claims lacked merit:

The Department had ample support for its conclusion that the continued integration of Claude into the Department’s information systems presented a statutorily covered national-security risk.

The statutory framework set forth in the Federal Acquisition Supply Chain Security Act authorizes agency heads to take “covered procurement actions” such as barring contracts or subcontract work with certain suppliers if doing so is necessary to reduce supply chain risk. Prior to taking such action, the agency head must make a written determination including that less intrusive measures are not reasonably available, notify the supplier, and provide an opportunity to respond unless an urgent national security interest justifies temporary delay.

In this case, Secretary of War Pete Hegseth issued a written determination finding substantial supply chain risk posed by Claude, emphasizing concerns that Anthropic might manipulate the AI’s design or operation in ways that could impair Department operations. Anthropic’s product includes embedded restrictions, or “constitutional training,” that prevent Claude from engaging in certain prohibited activities such as lethal autonomous warfare and domestic surveillance. These restrictions had resulted in Claude refusing legitimate government user requests in the past, causing operational uncertainty.

The court noted that Anthropic maintained these restrictions even after partial relaxation to accommodate government needs, and that a recent conflict arose when Anthropic’s team questioned the use of Claude in a sensitive military operation abroad, raising alarms in the Department and the prime contractor. The Department viewed an “any lawful use” contract term as essential, reflecting its AI-first strategy, and Anthropic’s refusal to assent to that term was central to the dispute.

On procedural grounds, the court affirmed its jurisdiction to review the Department’s actions. Anthropic timely filed a petition for review within 60 days of receiving notice of the exclusion. Although Anthropic later sought reconsideration, the court found jurisdiction secure either way, since the Secretary denied reconsideration and the petitions were consolidated.

On the merits, the court applied a deferential standard to the Department’s national security determinations, requiring only that the decisions be reasonable and reasonably explained. The court found that the record amply supported the Secretary’s conclusion that the exclusion was necessary to mitigate supply chain risk. It emphasized that Anthropic’s ability to enforce contractual usage restrictions via model training meant Claude might fail to perform as the Department expected, posing risks to critical defense systems.

The court also addressed Anthropic’s factual contention that the company could not access or modify the model after delivery. It clarified that while Anthropic could not control models post-delivery in real time, it could encode restrictions in new model versions supplied to contractors, a capability relevant to the supply chain risk assessment. The Secretary later confirmed that the determination did not rest on real-time control post-deployment.

Finally, the court rejected Anthropic’s claim that the statutory definition of supply chain risk did not cover a supplier’s refusal to permit certain uses of its product, reasoning that contractual manipulation of design or operation to deny lawful functions fell within the statute’s scope.

Editorial: Potential Impact

This decision signals judicial deference to executive determinations on national security risks posed by AI technology supply chains. The ruling underscores the government’s broad authority under the Federal Acquisition Supply Chain Security Act to exclude AI products from federal procurement where the supplier’s contractual usage restrictions frustrate government objectives or create uncertainty.

For the AI industry, the case demonstrates that companies providing AI models to government agencies may face pressure to accept broad “any lawful use” contractual terms to maintain access to defense contracts. This could influence how AI vendors structure usage policies and contract negotiations with government customers, especially concerning controversial applications like autonomous weapons and surveillance.

While the legal principles applied are based on a statute specific to federal procurement and national security, the decision may also inform broader debates about AI safety, control, and governmental reliance on privately developed AI technology. The court’s acceptance of the government’s factual concerns about model opacity and operational unpredictability suggests courts may be reluctant to second-guess executive assessments of AI risks tied to critical defense functions.

Source: Read the full decision on CourtListener

Court: Court of Appeals for the D.C. Circuit | Docket: 26-1049 | Date filed: 2026-09-25

This article was generated automatically from a public court record and is provided for informational purposes only. It is not legal advice. Refer to the linked source document for the authoritative text of the decision.