Pentagon Ends Use of Anthropic AI Over Safety Disputes
Listen
Vocabulary
Article
The US Department of Defence has stopped using Anthropic's Claude AI system, according to an official statement released in December. The Pentagon had deployed Claude since 2024 for research, analysis, and intelligence work, integrating it into Maven Smart System, a major data platform operated by Palantir. This marked the first time an advanced AI company gained access to classified government operations. However, the relationship deteriorated after the Pentagon requested that Anthropic remove safety guardrails from its tools. When Anthropic refused, Defence Secretary Pete Hegseth designated the company a supply-chain risk in February, citing national security concerns.
Integration and Removal Challenges
Anthropic's refusal centred on concerns about mass surveillance and autonomous weapons development. Rather than comply, the company sued the Trump administration, calling the designation unprecedented and unlawful. Meanwhile, the Pentagon signed contracts with competing firms including Google, xAI, and OpenAI, with OpenAI seeing wider adoption in recent months. Despite the February decision, Claude remained embedded throughout Maven until now. Lauren Kahn, a security analyst at Georgetown University, explained that removing Claude proved unexpectedly difficult because deep integration into existing systems makes extraction painful and complex.
Strategic Implications
This dispute reflects broader tensions between AI safety advocates and military innovation priorities. Anthropic's refusal to disable safety features represents a principled stand on responsible development, whereas the Pentagon viewed such restrictions as operational constraints. The delayed removal process reveals how difficult it is to extract AI systems once they become essential infrastructure. The outcome reshapes the Pentagon's AI landscape, consolidating its reliance on companies willing to meet military specifications while raising questions about whether safety and security objectives can coexist in defence applications.
Discussion Questions
- Why do you think the Pentagon found it difficult to remove Claude quickly from its systems, despite the February decision to stop using Anthropic products?
- Anthropic refused to remove safety guardrails because of concerns about mass surveillance and autonomous weapons. Do you agree that AI companies should maintain such protections even when military clients demand otherwise?
- How might the Pentagon's shift toward OpenAI, Google, and xAI tools affect the balance between AI capability and safety in military operations?
- The source notes this was unprecedented use of a supply-chain risk designation for a domestic AI company. What standards should governments apply when labelling private companies national security threats?
- Kahn observes that once AI systems become essential infrastructure, removing them is painful. What responsibility do organisations have when embedding external tools into critical systems?
Open the full interactive lesson with tap-to-hear on every word →