Pentagon Delivers Ultimatum to Anthropic: Remove AI Guardrails or Lose Military Contracts

Defense Secretary Pete Hegseth summoned Anthropic CEO Dario Amodei to the Pentagon on Tuesday morning for what senior Defense officials made clear was anything but a courtesy call. The meeting represents a breaking point between the U.S. military and the AI company behind Claude — the only artificial intelligence model currently operating inside the military’s classified systems — over whether a private technology company gets to set the rules for how its tools are used in war.

“Anthropic knows this is not a get-to-know-you meeting,” a senior Defense official told Axios ahead of Tuesday’s session. “This is not a friendly meeting. This is a sh*t-or-get-off-the-pot meeting.”

Pentagon officials say that weeks of negotiations have produced nothing and that the two sides enter the room on Tuesday from entirely different planets. Anthropic issued a statement calling the conversations “productive” and “in good faith.” Defense officials flatly disputed that characterization.

The core dispute is over what conditions Anthropic will place on military use of Claude. The company has sought formal assurances that its technology will not be used for mass surveillance of American citizens or to develop autonomous weapons capable of firing without a human in the decision chain. The Pentagon’s position, articulated by spokesman Sean Parnell last week, is that it wants to use Claude however it sees fit, provided the deployment does not violate the law. That gap — between Anthropic’s proactive ethical guardrails and the military’s operational flexibility demands — has proven unbridgeable through normal negotiations. Hence Tuesday’s meeting.

Hegseth arrives prepared to deliver an ultimatum. The Pentagon’s leverage is a potential designation of Anthropic as a “supply chain risk” — a label that carries devastating consequences. Under that designation, Anthropic’s existing government contracts would be voided, and every other company doing work with the Pentagon would be required to certify that it is not using Claude in any part of its workflow. Given how deeply embedded Claude has become across defense and intelligence operations, such a designation would create enormous disruption.

The catch, acknowledged by Defense officials themselves, is that offboarding Anthropic and replacing it with a comparable system would be a massive undertaking. No competing AI model currently available is considered as capable for sensitive defense and intelligence work. The Pentagon may be holding a loaded gun it would rather not fire.

The stakes climbed higher following reporting by The Wall Street Journal that Claude — accessed through Anthropic’s partnership with defense contractor Palantir Technologies — was used in the classified special operations mission that resulted in the capture of Venezuelan President Nicolás Maduro. That revelation made Anthropic the first known AI developer whose technology was deployed in a classified military operation.

Rather than cementing the relationship, the disclosure complicated it. Pentagon officials told Bloomberg News they became concerned after learning that Anthropic had questions about how its technology was used during the Maduro raid. Anthropic offered a different account of those communications, and the dispute over what was said and meant has added another layer of mistrust to an already strained relationship.

At the center of the culture clash is Amodei himself. A co-founder of Anthropic who previously worked at OpenAI, he has been among the most vocal voices in the AI industry warning about the risks of unchecked artificial intelligence development. He has built Anthropic’s brand around the concept of “constitutional AI” — embedding ethical constraints directly into how the model reasons and responds. That positioning has served Anthropic well in commercial markets and in public credibility debates, but it has put the company on a collision course with a Pentagon that views those constraints as obstacles rather than features.

“The problem with Dario is, with him, it’s ideological. We know who we’re dealing with,” a senior Defense official told Axios. Amodei has reportedly been resistant to the idea that his company’s safety frameworks are negotiable.

What makes this confrontation particularly significant is what it reveals about the larger question of AI governance in the national security space. For years, technology companies debated internally — and sometimes publicly — whether to work with the military at all. Google famously walked back its involvement in Project Maven in 2018 after employee protests. But Anthropic built its Pentagon relationship deliberately, partnering with Palantir to bring Claude into classified systems.

Now, having made that choice, it finds itself being told that access comes with conditions the company’s own ethics framework explicitly prohibits. The question before Amodei on Tuesday was not just whether to keep a contract. It was whether the government can compel a private AI developer to disable its own safety architecture as the price of doing business with the military.

There are reasonable arguments on both sides of the specific dispute. Autonomous weapons and mass surveillance of Americans represent genuinely consequential uses of technology, and a private company insisting on guardrails against both is not obviously wrong. At the same time, a military that cannot control how its own contractors’ tools function inside classified systems faces real operational problems.

What is harder to dismiss is the broader pattern: a federal government that has grown accustomed to treating tech companies as utilities, available for government use on government terms. The Anthropic confrontation may be the moment where that assumption gets tested against a company willing to say no — and where the consequences of that refusal get written into law, contract, and precedent. Tuesday’s meeting will not resolve those questions. But depending on what Amodei agreed to, or refused, it may define them.