Anthropic-Pentagon Dispute

Anthropic told a federal appeals court on April 22 that it cannot alter Claude after the software is installed inside classified Pentagon networks. (usnews.com) The filing is part of Anthropic’s fight against a Pentagon decision, made in early March, to label the company a national-security “supply chain risk.” That label bars defense contractors from using Claude on work tied to the Department of Defense. (cnbc.com) A three-judge panel on the U.S. Court of Appeals for the District of Columbia Circuit refused on April 8 to temporarily block the designation while the case proceeds. The court put the dispute on an expedited track and set oral argument for May 19. (politico.com) The technical point is simple: once an artificial-intelligence model is copied into a sealed government system, the vendor no longer has the same remote access it would have in a cloud service. Anthropic told the court that Claude instances running on classified networks are controlled by the government, not by Anthropic. (usnews.com) The legal fight grew out of a policy clash over military use, not a dispute over whether the Pentagon could use artificial intelligence at all. Anthropic said in March that negotiations broke down over two requested exceptions: mass domestic surveillance and fully autonomous weapons. (anthropic.com) Anthropic has also said it still supports many defense uses for Claude, including intelligence analysis, modeling and simulation, operational planning, and cyber operations. In July 2025, the Defense Department awarded the company a two-year prototype agreement with a $200 million ceiling. (anthropic.com, anthropic.com) The case now turns on who bears the risk when military customers run private-sector AI inside secure systems the vendor cannot directly control. That question reaches beyond Anthropic because the Pentagon is pushing more frontier AI into classified environments, where software oversight, contract terms, and liability all get harder to define. (apnews.com) Anthropic’s corporate structure is part of the argument around incentives. The company is a public benefit corporation, and its Long-Term Benefit Trust has authority that Anthropic says is meant to align the business with “the long-term benefit of humanity,” not only shareholder returns. (anthropic.com, anthropic.com) Anthropic has continued updating its own safety rules during the dispute. On February 24 it published version 3.0 of its Responsible Scaling Policy, and on April 2 it posted version 3.1, describing the framework it uses to assess catastrophic risks from more capable models. (anthropic.com, anthropic.com) The immediate question for the court is narrower than the politics around it: whether the Pentagon can keep treating Anthropic as a supply-chain threat even when Anthropic says the government, not the company, controls Claude inside classified systems. (usnews.com, politico.com)