Appeals Court Skeptical Anthropic Can Block US Supply-Risk Label

A federal appeals court appeared skeptical of Anthropic’s bid to block the Pentagon from declaring that the company poses a supply-chain risk to US national security, a move that led to a ban on government use of its artificial intelligence technology.

At a hearing Tuesday in Washington, two of the three judges on the panel peppered Anthropic’s attorney with questions about his claim that Defense Secretary Pete Hegseth made the declaration illegally in March, following a dispute over how the company’s Claude AI chatbot would be used by the military.

Judges Gregory G. Katsas and Neomi Rao — both appointed by President Donald Trump — questioned whether the court could legally second-guess Hegseth’s decision to reduce supply risk. They also suggested the unpredictable nature of AI tools and the government’s lack of trust in the vendor were sufficient justifications for the Defense Department to view them as a potential risk.

“It might not work in ways the government wants it to work,” Katsas said. “You still have this concern that the model is still unpredictable.”

“This is a fundamental black box,” added Rao.

Read More: Why Anthropic’s Mythos Is Sparking Global Alarm: Explainer

However, the third judge on the panel, Karen LeCraft Henderson, called Hegseth’s decision “a spectacular overreach by the government,” suggesting she’d side with Anthropic. “For the life of me, I don’t see any evidence of maliciousness” by the company, said Henderson, who was appointed by George H.W. Bush. She said the department had failed to support its claims.

The appeals court didn’t rule Tuesday, but the judges said in April that they’d work quickly to decide the case because Anthropic is likely to “suffer some irreparable harm” to its reputation and lost revenue.

The supply-risk declaration escalated a high-stakes dispute between the government and Anthropic.

The company wanted assurances its technology wouldn’t be used for mass surveillance of Americans or autonomous weapons deployment. The government rejected any restrictions, citing national security. A government court filing quotes Hegseth as having determined that Anthropic couldn’t be trusted and argued there was a risk the company could manipulate AI models to undermine the military’s use of classified information systems.

Anthropic has been fighting back on two fronts.

It asked the appeals court on March 9 to review the Pentagon’s declaration, calling it “arbitrary, capricious and an abuse of discretion.” The same day, Anthropic filed a separate complaint in San Francisco federal court challenging the ban. On March 27, a judge in that case blocked the government following through while the lawsuit proceeds. The Trump administration is appealing that decision.

At the hearing Tuesday, US Justice Department lawyer Sharon Swingle told the appellate panel the military was concerned Anthropic was effectively demanding “an operational veto” on lawful uses of its AI technology and there could be “red line” restrictions embedded in the software “that the government isn’t aware of.”

Anthropic attorney Kelly Dunbar disputed those claims, arguing the company had no capacity to alter its AI tools without the military’s permission nor a desire to second-guess operational decisions by the military.

Instead, Hegseth had a legal obligation to seek less intrusive measures than placing Anthropic in a risk designation typically reserved for companies from countries the US views as adversaries, Dunbar said. If the model doesn’t meet the government’s needs, it can just not buy it rather than imposing a declaration that is damaging to Anthropic’s reputation, he said.

Anthropic has argued it is being shut out for disagreeing with the administration. The company said the legal principles at stake affect every federal contractor whose views the government dislikes.

The case is Anthropic v. US Department of War, 26-01049, US Court of Appeals, District of Columbia Circuit (Washington).

© 2026 Bloomberg L.P. All rights reserved. Used with permission.