US Military AI Push Stumbles Over Anthropic Feud as China Advances

The Pentagon's drive to integrate advanced AI into weapons and cyber systems has collided with its own restrictions on Anthropic's models. In July the Air Force ordered contractors to remove Anthropic software by September 1 or risk losing Pentagon business. Weeks later the order was suspended, leaving contractors uncertain which rules apply.

The dispute traces to February, when Anthropic chief executive Dario Amodei told Defense Secretary Pete Hegseth the company would not supply models for domestic surveillance of Americans or for fully autonomous lethal weapons. Hegseth rejected the limits. The Pentagon labeled Anthropic a supply chain risk and barred its tools from most military networks, including the Maven intelligence-targeting system. The National Security Agency, however, retained limited experimental access to Anthropic's newer Mythos model for testing offensive cyber operations against foreign networks.

China's progress adds pressure. U.S. officials estimate Beijing is roughly six months behind on frontier models and years behind on advanced semiconductors, yet already fields cheaper alternatives that could target U.S. infrastructure. The administration has reversed export-control and access decisions multiple times this year while reviewing new models for biological-weapon and nuclear-system risks. No comparable restrictions apply to open-source models, including those from Chinese firms.