Anthropic Defies Pentagon Ultimatum Over Military AI Usage Restrictions

The high-stakes confrontation between Anthropic and the U.S. Department of Defense has reached a critical inflection point, marking a fundamental clash between Silicon Valley’s AI safety ethos and the Pentagon’s operational requirements. At the heart of the dispute is Anthropic’s refusal to lift usage restrictions that prevent its Claude models from being used for autonomous lethal targeting and domestic surveillance. For a company founded on the principle of constitutional AI, these safeguards are non-negotiable. However, for Defense Secretary Pete Hegseth, these restrictions represent a bottleneck to American technological superiority in an era of rapid AI militarization.

The ultimatum delivered to CEO Dario Amodei—to comply by Friday or face the invocation of the Defense Production Act (DPA)—is an unprecedented move against a software provider. Historically, the DPA has been used to compel the production of physical goods like steel or medical supplies. Applying it to force a change in a software’s usage policy or internal safety weights would set a massive legal precedent. It effectively signals that the U.S. government views AI models not just as commercial products, but as strategic national assets that can be seized or modified under the guise of national security. This move would essentially treat Anthropic's intellectual property as a public utility subject to executive override.

“

Companies like xAI and OpenAI have increasingly signaled a willingness to support defense initiatives, leaving Anthropic isolated in its safety-first stance.

This friction comes at a time when the Pentagon is aggressively diversifying its AI vendor pool to avoid vendor lock-in and ethical bottlenecks. While Anthropic once held a unique position as the sole Large Language Model (LLM) provider on certain classified networks, that monopoly has ended. The recent announcement that Elon Musk’s xAI has secured an agreement to deploy across classified networks suggests the Pentagon is willing to pivot toward providers with fewer ethical reservations. Companies like xAI and OpenAI have increasingly signaled a willingness to support defense initiatives, leaving Anthropic isolated in its safety-first stance. The presence of established defense contractors like Palantir in this ecosystem further highlights the shift toward a more integrated military-industrial-AI complex where speed and lethality often take precedence over alignment research.