The Pentagon isn't asking nicely anymore. Defense Secretary Pete Hegseth recently sent a tremor through the tech world by effectively telling Anthropic to get out of the way of national security. The message was blunt. If you build powerful AI on American soil, the American military expects to use it without a lecture on "safety" from a board of directors.
This isn't just another spat between a cabinet member and a CEO. It's a fundamental collision of two entirely different worlds. On one side, you have the "effective altruism" crowd at Anthropic, a company literally founded on the idea that AI is a terrifying existential threat that needs a "constitutional" leash. On the other side, you have a Defense Department that sees AI as the only thing standing between a controlled global order and a future dominated by the Chinese Communist Party. Expanding on this theme, you can find more in: Why the Green Party Victory in Manchester is a Disaster for Keir Starmer.
Hegseth’s warning to Anthropic isn't just about software. It's about sovereignty. He’s making it clear that the era of "Silicon Valley Neutrality" is dead.
The Anthropic Dilemma and the Safety Cult
Anthropic was born out of a schism. Dario and Daniela Amodei left OpenAI because they felt the company was becoming too commercial and playing too fast and loose with safety. They wanted to build a "helpful, harmless, and honest" AI. They even created a "Constitution" for their model, Claude, to ensure it wouldn't go rogue or help someone build a biological weapon. Analysts at The Guardian have provided expertise on this situation.
That sounds great in a vacuum. But Hegseth is looking at a different vacuum—the one left behind if the U.S. doesn't lead in tactical AI.
The military doesn't want "harmless" AI when it comes to electronic warfare, autonomous drone swarms, or predictive logistics in a South China Sea conflict. They want AI that wins. When Anthropic puts guardrails on its tech that prevent it from being used for "offensive" operations, they aren't just protecting the world from a robot uprising. They're potentially handicapping the very forces tasked with defending the country where Anthropic is headquartered.
Hegseth's point is simple. You can't claim to be an American company while refusing to help the American military maintain its edge against adversaries who have zero ethical qualms about AI safety.
Why the Pentagon is Losing Patience
For years, the Department of Defense (DoD) tried to play the "innovation partner" role. They set up offices in Silicon Valley. They spoke the language of venture capital. They tried to be the cool uncle with a big checkbook.
It didn't work. Google employees revolted over Project Maven. Microsoft faced internal protests over HoloLens. And Anthropic has remained one of the most hesitant players in the space.
But the geopolitical clock is ticking. China has integrated AI into its military strategy with a level of "civil-military fusion" that makes our current procurement process look like a bake sale. They don't have "safety boards" that can veto a general's request.
Hegseth understands that the next war won't be won by the side with the most tanks. It'll be won by the side with the fastest OODA loop (Observe, Orient, Decide, Act). AI is the engine of that loop. If Claude 3.5 or whatever comes next is the best reasoning engine on the planet, Hegseth believes it's a national asset, not just a private product.
The Problem With Constitutional AI in Combat
Anthropic’s "Constitutional AI" is a specific technical method. They train a model to follow a set of written principles. If those principles include "do not assist in violence," the model will refuse to help a targeting officer optimize a strike package.
Hegseth’s team sees this as a bug, not a feature.
Imagine a scenario where an officer needs to analyze satellite imagery to find a mobile missile launcher. If the AI flags that request as "potentially harmful" because it leads to a kinetic strike, the system is useless in a peer-to-peer conflict. Hegseth is essentially telling Anthropic that their internal ethics shouldn't override the chain of command.
The Legal and Financial Leverage
How does Hegseth actually move the needle here? He can't just walk into Anthropic’s offices and seize the servers. At least, not yet.
The leverage comes from three places:
- Cloud Providers: Anthropic relies on massive compute from Amazon and Google. Both companies have significant government contracts. Hegseth can lean on the infrastructure providers to ensure their tenants aren't obstructing national security interests.
- Export Controls: The government can make it very difficult for AI companies to operate globally if they aren't deemed "cooperative" on the home front.
- The Defense Production Act: This is the "nuclear option." If the President deems AI critical to national defense, the government can prioritize its orders and effectively direct how the technology is deployed.
Hegseth isn't invoking the DPA yet. He’s firing a warning shot. He’s telling the VCs and the founders that the "wait and see" approach to military integration is over.
A Culture War in the Clouds
There’s a deeper friction here that people often miss. Hegseth represents a "Peace Through Strength" worldview. Many at Anthropic represent a "Safety Through Restraint" worldview.
These two philosophies can't coexist when the stakes are this high. Hegseth’s background as a combat veteran makes him uniquely allergic to the "AI safety" jargon. To him, safety means your guys coming home and the other guys staying down. To a researcher in San Francisco, safety means the model doesn't say something "problematic" or help a hacker.
The disconnect is massive. Hegseth is basically saying, "I don't care if the AI is 'honest' to me, I want it to be 'lethal' to the enemy."
What Happens if Anthropic Refuses
If Anthropic keeps the gates closed, the DoD will just double down on companies like Palantir and Anduril. We're already seeing a massive shift in where defense dollars go.
But here’s the catch. Anthropic’s models are genuinely some of the best in the world. Claude’s ability to reason through complex, long-context data is exactly what the military needs for strategic planning. If the DoD is forced to use "lesser" AI because the "best" AI is too busy being ethical, that's a failure of national policy.
Hegseth’s pressure campaign is designed to prevent that two-tier system. He wants the A-team on the front lines.
Moving Beyond the Standoff
The reality is that Anthropic needs the government as much as the government needs them. The cost of training these models is skyrocketing into the billions. Private capital is great, but government-scale compute and long-term contracts are the bedrock of industry.
If you're following this space, stop looking at the press releases about "responsible AI." Start looking at the job boards at these companies. Look for the "Government Affairs" and "Public Sector" roles. That's where the real negotiation is happening.
Anthropic will likely find a way to create a "secure" or "defense-specific" version of their models. They'll call it a "specialized deployment" to save face with their safety-conscious employees. But make no mistake, Hegseth's bluntness worked. The wall between Big Tech's "safety" labs and the Pentagon’s war rooms is coming down.
Keep an eye on the next round of cloud infrastructure contracts. If Anthropic suddenly gets a massive boost in government-funded compute, you’ll know they’ve played ball. If they don't, expect more public call-outs from the Secretary of Defense.
Now is the time to watch how "AI ethics" survives a collision with the reality of 21st-century warfare. It’s usually the ethics that give way first.
