Judges Clash Over Pentagon’s Anthropic Ban

Artificial Intelligence & Machine Learning , Next-Generation Technologies & Secure Development , Standards, Regulations & Compliance Judges Clash Over Pentagon’s Anthropic Ban Appeals Court Weighs Pentagon Authority Over Frontier AI Providers Chris Riotta (@chrisriotta) • May 19, 2026     Share Post Share Credit Eligible Get Permission Image: JRdes/Shutterstock A majority of judges on a U.S. federal appeals court appeared disposed Tuesday to allowing Defense Secretary Pete Hegseth to bar artificial intelligence stalwart Anthropic from future military work for posing a risk to national security. See Also: Securing AI Workloads With Ubuntu Pro Oral argument held Tuesday in the U.S. Court of Appeals for the D.C. Circuit was Anthropic's latest salvo to get the designation overturned, a fight fought in multiple courts over how far the government can go to pressure frontier AI companies that resist unrestricted military use of their models. The supply-chain risk designation gave military agencies six months to stop using Anthropic technology but leaves non-defense government and commercial uses largely untouched. Anthropic attorney Kelly Dunbar argued before a three judge panel that the designation "defied congressionally mandated procedures, exceeded statutory limits and violated the Constitution." The Pentagon used a powerful national security tool to gain leverage in a dispute over how Claude could be used, she told Judges Gregory Katsas and Neomi Rao, both appointed by President Donald Trump, and Judge Karen Henderson, a George W. Bush appointee. "This court can’t review designation decisions" said Rao, who also noted that no national security risk designee has ever challenged the label in court. The best the court might be able to do, Rao said, is bar the Pentagon for again designating Anthropic a national security risk. The government is arguing that the issue is not speech, but whether defense officials can trust a model provider whose own usage limits would allegedly disrupt military operations. Henderson appeared skeptical of the government's position, questioning whether the record supports treating Anthropic as a supply-chain threat rather than a company engaged in a contract dispute with the military. "To me, this is just a spectacular overreach by the [Defense] Department," Henderson said. Rao pressed Anthropic on the limits of judicial review in national security procurement decisions, suggesting the Pentagon may have the latitude to assess risk when frontier AI systems are deployed in military contexts where failures could carry operational consequences. Katsas also focused on model opacity and the pace of AI development. Anthropic says its restrictions only apply to two high-level categories: systems that enable fully autonomous lethal weapons and mass domestic surveillance. It also has maintained that it cannot alter Claude once the model is deployed inside classified military environments. If the Pentagon does not want to use its technology, it can simply stop contracting with the company rather than attach a national security risk label with broader reputational consequences, the company has argued. Department of Justice attorney Sharon Swingle defended the designation as a necessary response to the Pentagon's loss of confidence in Anthropic. She argued Tuesday that the military could face catastrophic consequences if a model failed or became unavailable during active operations. "It's undisputed that the failure of the model in active military operations could have catastrophic national-security consequences and put service members' lives at risk," Swingle told the judges. The government has also argued that supply-chain risk authorities allow defense officials to act before a vulnerability becomes an operational failure. The legal fight has already produced conflicting results. The D.C. Circuit previously declined to pause the Pentagon's designation but expedited Anthropic's appeal, while a federal judge in San Francisco separately blocked the government from labeling Anthropic a supply-chain risk in a related case (see: Court Backs Pentagon Anthropic Ban - But the Fight Continues). The split has left Anthropic in a narrow and damaging position, according to analysts. The company remains barred from new Pentagon work - while continuing to serve non-defense government and commercial customers. The government is also wrestling with how to replace or retain access to frontier AI capabilities already embedded in sensitive environments (see: Federal Staffers Are Still Using Claude Despite Trump Orders). The Pentagon may be seeking a way out of the conflict, informing the circuit judges earlier this month that it has decided to grant Anthropic's grant for a reconsideration of the designation. In a court filing, Defense attorneys said Anthropic had requested a reconsideration, but didn't send it to the correct official email account. "Although the email was not submitted according to the specified procedures, the department intends to treat it as a timely request for reconsideration," government attorneys informed the court. The dispute has also become more complicated as the administration has pursued broader frontier AI deployment across government networks. The Pentagon recently announced a series of agreements with multiple AI companies, including OpenAI, Google, Nvidia, Microsoft and Amazon Web Services - a move Anthropic supporters say undermines the government's claim that pausing the designation would threaten military readiness. Outside national security officials have warned the designation could chill defense innovation by signaling that commercial AI companies may face punitive consequences if they refuse to abandon internal safety policies. Former defense and intelligence officials told the court in amicus filings that the designation appeared pretextual and risked undermining trust between the Pentagon and the private firms it increasingly depends on for emerging technology. Anthropic has said it supports national security applications of Claude and has previously worked with the Pentagon on intelligence analysis, modeling and simulation, operational planning and cyber operations.