Appeals Court Lets the Pentagon Blacklist Anthropic, 2 to 1
The safeguards are the problem. That is the core of what the U.S. Court of Appeals for the D.C. Circuit said on September 25 when it upheld, 2 to 1, the Pentagon's designation of Anthropic as a supply chain risk. The majority found the Defense Department had enough grounds to conclude that Claude's built-in restrictions, plus an unresolved contract dispute, could compromise military reliability. Anthropic's First Amendment and procedural claims were rejected.
Quick background. Anthropic signed a $200 million Pentagon contract in July 2025. Talks over putting Claude on the GenAI.mil platform broke down when the department wanted use for all lawful purposes and Anthropic wanted assurance of no fully autonomous weapons and no domestic mass surveillance. In March the department applied the supply chain risk label, which in practice locks Anthropic out of defense contracts and the contractor supply chain beneath them.
Judge Karen LeCraft Henderson's dissent is the sentence to remember: the law does not treat a contractor's honest and upfront enforcement of restrictions as grounds for blacklisting. That is exactly the question. The majority effectively said a vendor that tells you in advance what its model will refuse to do is a reliability risk for that reason.
It is not over. A separate district court ruling in August, from Judge Rita Lin, went the other way, finding the government wanted to make a public example of Anthropic for criticizing it, and that ruling still stands. Anthropic says it is considering all options including further review. Two federal courts are now pointing in opposite directions on the same designation.
For anyone building agents on Claude, the practical read is narrow but real. Usage policies are now a procurement variable, not just an ethics page. If published refusals can be treated as a supply chain defect by the largest buyer in the country, every lab's acceptable use policy just became part of its sales pitch or its liability, depending on the customer. Coverage: CNBC, ABC News and The Washington Post.
← Back to all articles
Quick background. Anthropic signed a $200 million Pentagon contract in July 2025. Talks over putting Claude on the GenAI.mil platform broke down when the department wanted use for all lawful purposes and Anthropic wanted assurance of no fully autonomous weapons and no domestic mass surveillance. In March the department applied the supply chain risk label, which in practice locks Anthropic out of defense contracts and the contractor supply chain beneath them.
Judge Karen LeCraft Henderson's dissent is the sentence to remember: the law does not treat a contractor's honest and upfront enforcement of restrictions as grounds for blacklisting. That is exactly the question. The majority effectively said a vendor that tells you in advance what its model will refuse to do is a reliability risk for that reason.
It is not over. A separate district court ruling in August, from Judge Rita Lin, went the other way, finding the government wanted to make a public example of Anthropic for criticizing it, and that ruling still stands. Anthropic says it is considering all options including further review. Two federal courts are now pointing in opposite directions on the same designation.
For anyone building agents on Claude, the practical read is narrow but real. Usage policies are now a procurement variable, not just an ethics page. If published refusals can be treated as a supply chain defect by the largest buyer in the country, every lab's acceptable use policy just became part of its sales pitch or its liability, depending on the customer. Coverage: CNBC, ABC News and The Washington Post.
Comments