Anthropic has told a federal appeals court in Washington that once its AI models are deployed inside the Pentagon's environment, the company has no visibility and no technical means to control or shut them down. There is no "kill switch." The company also noted that the Pentagon had the opportunity to test the models before deployment.
The filing is the latest development in a dispute over the Pentagon's "supply chain risk" designation. In March, the Pentagon labeled Anthropic a supply chain risk, arguing that the company improperly interfered with how its technology is used in sensitive military operations. At the heart of the fight is Anthropic's usage policy, which prohibits using Claude for autonomous weapons or mass surveillance. The Pentagon has called those restrictions a "smokescreen."
The litigation has produced a split between two courts: a Washington court rejected Anthropic's request to pause the supply chain risk label, while a California court granted it. The practical effect is that Anthropic cannot bid on new Pentagon contracts but can continue serving other government agencies. Meanwhile, the Trump administration is pushing to deploy Anthropic's new model, Mythos, across federal agencies. Agency heads are studying how to use Mythos to defend against cyberattacks — a stance that contradicts the Pentagon's claim that Anthropic poses a national security risk. The next hearing is scheduled for May 19.