Pentagon and Anthropic Clash Over AI Safety
The Department of Defense and AI firm Anthropic are in a dispute over the use of the company's Claude large language models on classified systems, placing a $200 million defense contract under review. The Pentagon is reportedly considering designating Anthropic a “supply chain risk” due to the company’s insistence on strict safety guardrails and limited military use cases. The conflict is being watched as a test of how AI vendors align with the DoD's June 2024 AI governance framework.
- The "supply chain risk" designation the Pentagon is considering for Anthropic is a measure rarely used against domestic firms and is more commonly applied to entities linked to foreign adversaries. If implemented, it would require all companies doing business with the military to certify they do not use Anthropic's Claude AI, impacting contractors far beyond the $200 million agreement under review. - The core of the dispute lies in Anthropic's specific safety policies, which prohibit its AI from being used to develop weapons that can fire without human intervention or to conduct mass surveillance on American citizens. The Pentagon is demanding use for "all lawful purposes," a position publicly reinforced by Defense Secretary Pete Hegseth. - Unlike Anthropic, competitors such as Google and OpenAI have reportedly agreed to the Pentagon's terms for all lawful uses of their models, at least on unclassified systems. However, Anthropic's Claude is currently the only large language model of its kind operating on the military's classified networks, a deployment that occurred through a partnership with defense tech firm Palantir. - The contract is a two-year prototype "other transaction agreement" (OTA) managed by the DoD's Chief Digital and Artificial Intelligence Office (CDAO), with a ceiling of $200 million. However, as of early February 2026, only $1.99 million of the total ceiling had been formally obligated. - The conflict with Anthropic stands in contrast to the deepening relationships between the Pentagon and other tech startups like Anduril and Scale AI. Anduril has recently won a $642 million counter-drone contract with the Marine Corps and a $100 million deal with the CDAO, while Scale AI has a $100 million agreement to deploy its AI platform on classified DoD networks. - There is a precedent for tech company pushback on Pentagon AI projects, most notably Google's 2018 withdrawal from Project Maven, an initiative to use AI to analyze drone surveillance footage, following employee protests. Project Maven has since continued with other contractors, including Palantir, and is now managed by the National Geospatial-Intelligence Agency.
