The Pentagon Just Tried to Kill an AI Company

Friday evening, the Trump administration didn't just end a contract dispute with an AI company. They attempted what legal observers are calling "corporate murder."

The sequence of events was brutal and compressed. At 3:47 PM Eastern, President Trump posted to Truth Social directing every federal agency to immediately cease using Anthropic's technology. By evening, Defense Secretary Pete Hegseth had declared Anthropic a supply chain risk—a designation previously reserved for foreign adversaries like Huawei—and announced that no Pentagon contractor could do any business with them. Not just on defense contracts. Any business, period.

The immediate legal problem: Amazon Web Services, Google Cloud, and Microsoft Azure all host Anthropic's models. All three companies have extensive Pentagon contracts. If you take Hegseth's directive literally, Anthropic's cloud infrastructure partners would need to choose between the company and the entire defense market.

"This is simply attempted corporate murder," wrote Dean Ball, who helped draft Trump's AI policy. "I could not possibly recommend investing in American AI to any investor."

The dispute ostensibly centers on two restrictions in Anthropic's terms of service: their AI model Claude cannot be used for mass domestic surveillance of Americans or for fully autonomous weapons systems. Anthropic CEO Dario Amodei argues these aren't arbitrary moral posturing—Claude isn't reliable enough for autonomous weaponry, and AI surveillance lacks adequate legal safeguards.

The Pentagon's position, articulated through increasingly heated public statements, is that private companies don't get to set constraints on how the government uses technology it purchases. Assistant Secretary Sean Parnell insisted the Department "has no interest in using AI to conduct mass surveillance of Americans, which is illegal. Nor do we want to use AI to develop autonomous weapons that operate without human involvement."

Which raises the obvious question: if the Pentagon has no intention of using AI for those purposes, why does accepting terms that prohibit them constitute a red line?

The answer appears to be pure principle—specifically, the principle that civilian companies shouldn't constrain military decision-making. Hegseth's statement was explicit: "The terms of service of Anthropic's defective altruism will never outweigh the safety, the readiness, or the lives of American troops on the battlefield. Their true objective is unmistakable: to seize veto power over the operational decisions of the United States military."

What makes this fascinating is how cleanly it separates different kinds of objections. The public debate isn't really about whether AI surveillance or autonomous weapons are good ideas—though plenty of people have strong views on that. It's about where authority over these technologies should rest.

Consider the positions now circulating:

Anthropic is right—some red lines shouldn't be crossed
Anthropic might be morally correct, but governments can't be constrained by vendors
Private companies shouldn't set government policy, regardless of whether their position is correct
The government should be free to decline working with vendors, but perhaps shouldn't try to destroy them economically
Making an example of Anthropic sends a useful message to other AI CEOs with ideas

The positions aren't perfectly nested—you can believe Anthropic is wrong about the technology and still think the supply chain risk designation is government overreach. You can support the military having full access to AI tools and still find the "effective immediately" declaration legally suspect.

Legal experts immediately flagged the process questions. Under 10 USC 3252, a supply chain risk designation requires completing a risk assessment, making a written determination that the designation is necessary for national security with no less intrusive alternative, and notifying Congress. "It's hard to believe they fulfilled, e.g., the congressional notice requirement in the time between 5:00 PM Eastern and Hegseth tweeting," wrote senior research fellow Charlie Bullock.

Senator Tom Tillis captured the surreal quality of watching this unfold on social media: "Why the hell are we having this discussion in public? Why isn't this occurring in a boardroom or in the secretary's office?"

Meanwhile, OpenAI was having very different conversations with the Pentagon. While Sam Altman publicly expressed support for Anthropic's safety principles—"We've long believed that AI should not be used for mass surveillance or autonomous lethal weapons"—his company was simultaneously negotiating its own deal. By Friday evening, OpenAI announced an agreement to deploy models in classified networks with language affirming those same principles.

The key difference: OpenAI agreed to build their own "safety stack" of technical controls rather than restricting the Pentagon's usage in the terms of service. If the model refuses a task, the government won't force OpenAI to make it comply. It's the same practical outcome with a different locus of control.

Altman framed this as a template: "We are asking the DoD to offer these same terms to all AI companies, which in our opinion, we think everyone should be willing to accept."

Anthropic, notably, had not received any formal communication from the Defense Department or White House beyond what everyone else saw on social media. Their response statement focused on reassuring customers that the supply chain designation, if formalized, would only affect Pentagon contract work—not commercial or individual access to Claude.

This is technically correct but practically optimistic. Anyone who's studied Operation Chokepoint—the Obama-era effort to cut off banking access to disfavored industries—knows that formal restrictions are less important than implied pressure. When the government signals a company is radioactive, other companies tend to keep their distance regardless of legal technicalities.

What's notable about the public reaction is how it refuses to map onto standard partisan lines. Eric Voorhees, founder of Venice AI and a libertarian willing to criticize both sides, put it bluntly: "Anthropic is definitely woke and lefty, but their refusal to permit Washington to use their tech to carry out warrantless mass surveillance of Americans is eminently based."

The "but China" responses were predictable but revealing. The argument goes: Chinese AI companies don't impose restrictions on their government, so American companies that do are effectively handing technological advantage to an adversary that won't hesitate to use it.

This argument has real force—until you remember that Anthropic has been more vocal than most AI companies about restricting advanced technology exports to China. Amodei wrote in his statement: "I believe deeply in the existential importance of using AI to defend the United States and other democracies and to defeat our autocratic adversaries." The company isn't pacifist. They're making a specific claim about specific use cases.

The question of who controls AI—the title of the video covering all this—doesn't have an obvious answer. The technology is being developed by private companies with their own governance philosophies, funded partly by government contracts, regulated by agencies still figuring out what regulation should look like, and deployed in contexts where the consequences of failure range from annoying to catastrophic.

Anthropic refused to bend. The government responded by attempting to cut off their oxygen supply. OpenAI found a formulation both sides could accept. And the rest of the AI industry is now watching to see whether having principles about your technology's use is compatible with building a viable American AI company.

That's not a hypothetical question anymore.

—Marcus Chen-Ramirez