The Pentagon is trying to bully Anthropic into loosening Claude’s guardrails, claiming that if it can’t use it to spy on people and automatically kill people without human involvement, commit war crimes, it is not good for anyone.
Pentagon warns Anthropic it will “pay a price” unless it does as it is told.
The US defence establishment is reviewing its relationship with Anthropic after months of tense negotiations over how the military can use the firm’s Claude model, which is currently the only AI system running inside classified Pentagon networks.
Chief Pentagon spokesman Sean Parnell used Axios to fire a warning shot, saying: “The Department of War’s relationship with Anthropic is being reviewed. Our nation requires that our partners be willing to help our warfighters win in any fight. Ultimately, this is about our troops and the safety of the American people.”
Behind the flag-waving is a concrete contractual fight. The Pentagon wants Claude and other frontier models available for “all lawful purposes”, which sounds tidy until you remember how elastic “lawful” gets when you talk about war.
Anthropic is reportedly willing to loosen its current terms of use but wants to block two specific outcomes: tools used to spy on Americans at scale and the development of weapons that fire without human decision-making.
Defence officials argue that those restrictions are too tight and too vague to be workable, claiming that the real world is full of grey areas that would make Claude difficult to deploy across the kinds of tasks modern militaries undertake.
The Pentagon is pushing the same “all lawful purposes” line in talks with three other big AI labs, OpenAI, Google and xAI, and officials appear confident they will sign up, even if the details are still being fought over.
People close to the discussions say senior defence officials have long been frustrated with Anthropic and have seized the opportunity to turn the dispute into a public spectacle.
Anthropic has sought to be transparent publicly. A spokesperson said: “We are having productive conversations, in good faith, with DoW on how to continue that work and get these new and complex issues right.”
The company has been reminding everyone that Claude was the first model deployed on classified networks, which makes the Pentagon’s current posture appear to threaten a tool it already relies on.
Claude is reportedly deeply embedded across military workflows and was used during a January raid involving Venezuela’s Nicolás Maduro, a detail that can turn “experimental” software into “operationally essential” kit.
The Pentagon’s main threat is bureaucratic. It may label Anthropic a “supply chain risk”, a designation that would force the sprawling universe of Pentagon vendors to certify they do not use Claude in their own internal workflows.
Given how widely Anthropic is used across corporate America, that is a nasty lever, because it makes Claude radioactive inside any company that wants to sell anything to the Pentagon.
The contract at the centre of the dispute is valued at up to $200 million, which is pocket change next to Anthropic’s reported $14 billion in annual revenue, but the supply chain blast radius could be far more damaging than the cancelled deal.
There is a practical problem for the Pentagon if it follows through. A senior administration official has said rival models are “just behind” for specialised government applications, which means a forced switch could lead to a messy scramble within sensitive systems.
Civil liberties concerns sit in the background of all this. Existing US mass surveillance law was not written with frontier AI in mind, and critics warn that plugging AI into collection and analysis pipelines could supercharge the government’s ability to target civilians.
That is the real reason Anthropic is trying to hold the line on surveillance and autonomous weapons, even as the Pentagon waves the flag and insists it just wants flexibility for “lawful” missions.
For now, the situation looks less like a principled national security debate and more like a procurement brawl where one side wants fewer rules and the other wants to avoid being the company that helped automate the wrong kind of headline.
