There are phrases in the world of technology policy that seem taken from a science fiction dystopia but are real. That of February 27, 2026 is one of them: Pete Hegseth, Secretary of Defense of the United States, publicly declared that Anthropic—the artificial intelligence company founded by Dario Amodei and several former OpenAI researchers—constituted a "risk to the supply chain."
The name "supply chain risk" is a technical-legal term that the Department of Defense normally applies to companies with direct connections to foreign adversaries, suppliers with links to hostile governments, or entities that could compromise the national security of the United States through their products. Huawei is a supply chain risk. A Californian startup whose founders come from DeepMind and OpenAI, which has received billions in investment from Google and Amazon, and which competes directly with American companies like OpenAI in the same market... is not exactly the usual profile.
Why it really happened
To understand why Anthropic ended that label, you have to go back to January 2026. Secretary Hegseth signed a memorandum that established a new condition for all AI contracts with the Pentagon: companies had to grant the use of their models for "any lawful government purpose," without additional restrictions imposed by the supplier itself.
Anthropic had two red lines that Dario Amodei had publicly stated it was unwilling to cross: that its technology be used in mass surveillance of American citizens, and that it operate fully autonomous weapons systems without direct human supervision. These were not improvised positions: they were integrated into the company's terms of service, its acceptable use principles, and its internal constitution.
"We cannot, in good conscience, grant your request." — Anthropic's official statement to the Pentagon, refusing to remove ethical safeguards from its AI models
When the Pentagon demanded that Anthropic remove those restrictions in order to sign the contract, the company said no. Not as a negotiation tactic, but as a definitive position. "We cannot, in good conscience, agree to your request," they said in an official statement. The government's response was swift: Trump announced that all federal agencies must stop using Anthropic products, and Hegseth applied supply chain risk classification. On the same day, OpenAI signed its contract with the Pentagon.
The irony of classification
The irony of calling Anthropic a "supply chain risk" is that the company was founded precisely by people who left OpenAI because they believed that that company was not careful enough with AI security. Its development methodology, called Constitutional AI, was specifically designed to create models that follow ethical principles more robustly than alternative approaches. Its most important investors include Google and Amazon — hardly adversaries of the American state.
What the classification communicates, in plain language, is this: the United States government considers a company that refuses to give up control of its technology for uses it considers ethically unacceptable to be a national security problem. Not because of his foreign connections. Not because of their data practices. But for having principles that do not fit the requirements of the government client.
The two red lines that Anthropic did not cross
- Domestic mass surveillance: Anthropic refused to allow the use of Claude to surveil US citizens without judicial oversight
- Autonomous weapons without human control: Anthropic refused to give up control on lethal weapons systems that operate without direct human decision
- Both restrictions have been part of the company's public terms of service since its founding
- Anthropic announced it would legally challenge the "supply chain risk" classification
Anthropic's legal response
Anthropic did not remain silent. The company announced that it planned to legally challenge the supply chain risk designation, pointing out that this classification is normally applied to companies with direct connections to foreign adversaries and that applying it to an American startup with Anthropic's track record was a distortion of the legal mechanism. In practical terms, the classification could extend beyond the Pentagon and affect other government contracts, posing an existential threat to a significant segment of the company's business.
The episode also had a consequence that few analysts expected: Pentagon officials themselves began to question the decision internally. According to Fortune reports, there was a "moment of alarm" when defense directors realized how much they depended on Anthropic's work in areas of research that they couldn't easily replicate with other vendors. Anthropic's classification as a risk was, in that sense, a shot in the foot: the Pentagon needed Anthropic's technology but had made it an ineligible contractor.
The precedent and its global implications
Beyond the specific case of Anthropic, the episode sets a worrying precedent for the entire AI industry. Communicate to companies in the sector that maintaining ethical safeguards that conflict with the requirements of the world's largest government client can have severe commercial consequences. It is a structural incentive for AI companies to gradually abandon ethical restrictions that make contracting with governments difficult.
In Europe, reading of the episode was received with a mixture of alarm and satisfaction: alarm because it confirms that the militarization of AI is advancing faster than any regulatory framework can keep up, and satisfaction because it validates the European approach of trying to establish legal limits before market dynamics make them impossible to enforce.
The story of Anthropic and the Pentagon is not over. The legal challenge is still ongoing, the supply chain risk classification has practical effects on the contracts in force, and the debate over whether AI companies can maintain ethical principles when the client is the most powerful government in the world is one of the most important in the coming years. For now, Anthropic still has not crossed its red lines. How long he can stay in that position is an open question.
Share this article
