Pentagon threatens Anthropic over AI access

Anthropic is refusing a Pentagon demand to provide unrestricted access to its AI model, setting up a rare test of how far the US government is willing to go to override private “guardrails” with national-security authority. According to Axios, cited by TechCrunch, Defense Secretary Pete Hegseth told Anthropic CEO Dario Amodei on Tuesday that the Department of Defense could label the company a “supply chain risk” or seek to invoke the Defense Production Act (DPA) to compel a tailored military version of the system.

The confrontation is less about a single contract than about who gets to write the operating rules for a general-purpose technology that is quickly becoming infrastructure. Anthropic has said it does not want its models used for mass surveillance of Americans or for fully autonomous weapons. Pentagon officials counter that military use should be constrained by US law and constitutional limits—not by a vendor’s usage policy. In practice, that is a dispute over enforcement: constitutional limits are litigated after the fact, while vendor restrictions can block capabilities at the point of use.

The leverage on Washington is that Anthropic is, by several accounts, the only frontier AI lab currently cleared for classified DoD access. TechCrunch notes that the Pentagon has reportedly reached a deal to use xAI’s Grok in classified systems, but the broader problem is redundancy. Dean Ball of the Foundation for American Innovation told TechCrunch the DoD appears exposed to a single-vendor dependency that conflicts with a late-Biden National Security Memorandum directing agencies to avoid reliance on one classified-ready frontier AI system. If the government cannot swap suppliers quickly, it has an incentive to turn procurement tools into coercive tools.

Using the DPA in a fight over model restrictions would also expand a law that was widely used during the COVID-19 pandemic to force production of physical goods like ventilators and masks. Software is different: the “production” being compelled is policy compliance, model tuning, and access controls—areas where the contractor’s product is inseparable from its governance. A DPA-driven outcome would not just deliver a capability to one customer; it would likely create a template for audits, certifications, and carve-outs that other agencies and contractors would treat as default.

The standoff is unfolding amid open political signaling. TechCrunch reports that figures in the administration, including AI czar David Sacks, have attacked Anthropic’s safety policies as “woke.” Ball argues that weaponizing designations like “supply chain risk” against a domestic supplier would make the US appear less predictable to investors and executives.

Anthropic has until Friday evening to change course, according to Axios. One of the government’s most powerful emergency authorities is now being discussed as a way to rewrite a private company’s terms of service.