Pentagon Considers Ending Anthropic AI Partnership

- The two-year prototype agreement, with a ceiling of $200 million, was awarded in July 2025 by the Pentagon's Chief Digital and Artificial Intelligence Office (CDAO). This office was established in 2022 to accelerate the adoption of data and AI across the Department of Defense. - The core of the dispute centers on Anthropic's specific "hard limits" against using its models for fully autonomous weapons systems and mass domestic surveillance. The Pentagon, in contrast, is pushing for an "all lawful purposes" standard in its contracts with AI vendors. - Anthropic's stance is guided by its public "Responsible Scaling Policy," which outlines AI Safety Levels (ASLs) that trigger stricter controls as model capabilities increase to manage catastrophic risks. The company's acceptable use policy explicitly prohibits applications in "military and warfare" and "weapons development." - The CDAO has awarded similar contracts of up to $200 million each to Google, OpenAI, and xAI, creating a competitive environment where other major labs appear more flexible to the Pentagon's requirements. - This conflict reflects a broader trend among major AI companies shifting their stance on military work. OpenAI, for instance, removed language from its usage policy in early 2024 that had banned "military and warfare" applications, though it still prohibits its tech from being used to "develop or use weapons." Google also revised its AI Principles in 2025, removing a previous ban on developing AI for weapons that was instituted after employee protests over its involvement in Project Maven. - Project Maven, a key Pentagon AI initiative started in 2017, provides context for the military's goals. It uses machine learning to analyze data from drones and other sources to identify potential targets for human verification, and has been used in operations in the Middle East and to support Ukraine. - The contract with Anthropic was intended to have the company develop working prototypes fine-tuned on DoD data and collaborate with defense experts to mitigate potential adversarial uses of AI. - The Pentagon has reportedly considered designating Anthropic a "supply chain risk" if an agreement cannot be reached. This would effectively force any company wanting to work with the DoD to sever ties with Anthropic.