Regulating Against Safety

How the Pentagon's impending ultimatum to Anthropic weaponizes the law to outlaw AI ethics.

Feb 27, 2026

The Hammer and the Guardrail

On March 27, 2020, Donald Trump tweeted two words about General Motors: “Invoke P.”

He meant the Defense Production Act. A Korean War statute, passed in September 1950, that gave the president authority to compel private American companies to produce materials necessary for national defense. Trump had been resisting using it for weeks. The US Chamber of Commerce told him not to. He’d been saying publicly that companies were volunteering their capacity and didn’t need threats. “Frankly, they don’t need someone to walk over there with a hammer and say do it.”

Then negotiations with GM over ventilator pricing hit a snag, and Trump walked over there with the hammer.

Last week, Pete Hegseth used the same hammer on Anthropic.

a sign with a question mark and a question mark drawn on it — Photo by Nahrizul Kadri on Unsplash

The Defense Secretary threatened to cancel a $200 million Pentagon contract, declare Anthropic a “supply chain risk,” and invoke the Defense Production Act to force the company to remove its AI safety guardrails and grant the military full access to its models. Dario Amodei — who had warned publicly against using AI for mass surveillance and autonomous weapons — has until 5:01 PM Friday to comply.

Everyone is covering this as a tech story, or a civil liberties story, or a story about Dario Amodei’s backbone.

It’s bigger than all of those. And it has been 76 years in the making.

Transparency Cascade Press is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.

The Paradigm Shift Nobody Is Naming

There’s a distinction that most of the coverage is missing, and it’s the most important thing about this story.

For the last decade, the debate about AI and government has been about deregulation — about whether Washington would get out of the way and let companies build without restriction. Safety advocates wanted guardrails. Industry sometimes resisted. The question was whether the government would require safety or simply permit recklessness.

What Hegseth did last week is categorically different. The Pentagon is not deregulating. It is regulating against safety — using the coercive power of the state to mandate the removal of protections that a private company voluntarily chose to build.

AI safety protocols are no longer just being ignored by the government. They are being officially treated as obstacles to national security that must be dismantled by statutory force.

That is a complete paradigm shift. And once the precedent is established, it cannot easily be undone.

The Template (1917)

Before the Defense Production Act, there was the War Industries Board.

Bernard Baruch ran it during World War I with “dollar-a-year men” — corporate executives who lent their expertise to the government for nominal pay, then returned to their industries when the war ended. The model was explicit: government and industry coordinate for war, then part ways. The emergency powers were temporary.

That assumption would not survive the next war.

The Permanent Emergency (1950)

On April 7, 1950, Paul Nitze’s NSC-68 made the war economy permanent. George Kennan opposed it. He had designed containment as a political strategy; Nitze remade it as a military one, arguing that the Soviet threat required sustained defense spending at a scale that would reshape the American economy.

Korea locked it in before Congress could debate the implications.

Five months later, on September 8, 1950, Congress passed the Defense Production Act. It gave the president authority to prioritize, allocate, and compel production from private companies in the interest of national defense. Unlike the War Industries Board, it didn’t disband when the shooting stopped. It stayed on the books. Congress has reauthorized it, with expansions, ever since.

Eleven years later, on January 17, 1961, Eisenhower gave his farewell address. He had originally written “military-industrial-congressional complex.” He deleted “congressional” before delivery — a president, he felt, shouldn’t criticize Congress on his way out.

The most important word in the most important warning ever issued about American power was cut before anyone heard it.

🔗 View the DPA timeline: from Korean War to Hegseth

The Limit (1952)

Two years after the DPA passed, Truman tested it — and lost.

Facing a nationwide steel strike during the Korean War, he seized the steel mills by executive order, citing his Commander-in-Chief authority. He bypassed the DPA’s own seizure mechanisms, which required Congressional involvement he didn’t want.

The Supreme Court ruled 6-3 in Youngstown Sheet & Tube Co. v. Sawyer that he had no authority to do so. Justice Robert Jackson’s concurrence became the governing framework for executive power ever since: a president acting against the implied will of Congress operates at the “lowest ebb” of constitutional authority.

Here is what matters for 2026: Truman lost Youngstown because he acted outside the statute. Hegseth is acting inside it. The DPA expressly authorizes the executive to compel private companies to serve national defense purposes. That’s Jackson’s first category — maximum presidential power, backed by Congressional authorization.

The judicial limit that stopped Truman doesn’t apply when you use the law Congress actually passed.

The Normalization (2020-2022)

For seventy years, presidents used the DPA carefully and narrowly. The Pentagon placed roughly 300,000 DPA orders per year for military equipment. It was administrative infrastructure, not a headline.

Then came COVID. Trump invoked it against GM. The Chamber of Commerce’s objections were overridden. Voluntary cooperation became legally irrelevant once the hammer came out.

Biden went further. On June 6, 2022, with his legislative climate agenda stalled in Congress, he invoked the DPA five times in a single day — for solar panels, heat pumps, building insulation, electric grid transformers, hydrogen electrolyzers. Republicans were furious. Senator Pat Toomey called it an abuse of a law designed for Korean War defense mobilization.

They were right about the precedent. They were wrong about whose goals it would eventually serve.

The party that howled about Biden’s solar panels built the legal architecture for Hegseth’s ultimatum.

🔗 View DPA invocations 2020-2026: from ventilators to AI

The Inversion (2026)

The Defense Production Act was written to make things.

Make ventilators. Make tanks. Make solar panels. Make vaccines. The entire 76-year history of the statute is a history of the government compelling more — more output, more supply, more capacity.

Hegseth used it to compel less.

Not less production. Less safety. Less caution. Less human oversight of autonomous weapons and mass surveillance systems.

The power to compel production became the power to compel recklessness.

There is no entry in the 76-year history of the Defense Production Act for this use. Every prior invocation expanded what a company would do or make. The Hegseth ultimatum is the first to demand that a company remove something it had committed to.

And to understand what’s actually being demanded, you have to look at what Anthropic’s guardrails actually were.

What the Guardrails Actually Protected

Anthropic’s safety commitments weren’t abstract corporate ethics theater. They were specific, targeted, and — as it turns out — directly in the Pentagon’s way.

Autonomous lethal force. For decades, DoD Directive 3000.09 has required “appropriate levels of human judgment” over lethal force — what arms control advocates internationally call “meaningful human control.” Anthropic’s guardrails went further, prohibiting fully autonomous weapons outright — a stricter standard than the Pentagon’s own doctrine. By forcing their removal, the Pentagon is not just asking Anthropic to drop a corporate policy. It’s using a private company’s compliance to route around its own directive — signaling a shift toward systems that can target and fire without human intervention.

Mass domestic surveillance. The Fourth Amendment protects against unreasonable search — but it was written for a world where surveillance required human beings and physical resources, which created natural friction. Advanced AI breaks that framework entirely. It enables surveillance at a scale, speed, and cost that eliminates the friction that traditionally required warrants. Anthropic’s guardrails against mass domestic surveillance were a check on a capability that the Fourth Amendment was never designed to address. Hegseth’s ultimatum removes that check. The legal framework hasn’t caught up. The capability is about to be unleashed anyway. (The surveillance infrastructure already exists. Palantir’s ImmigrationOS — which I documented last week — already lets ICE agents draw shapes on a map and target everyone inside. Removing AI safety guardrails gives that system a brain.)

The “All Lawful Purposes” mandate. Hegseth’s directive demands that AI models be available for “all lawful purposes” without what he calls “ideological constraints.” Read that carefully. The Pentagon is telling tech companies that corporate ethics policies are incompatible with defense contracts. If you want to be a US defense contractor — and in the AI era, that market is enormous — you don’t get to decide how your technology is used, even when the use involves automated lethal force or warrantless mass surveillance of American citizens.

This is not deregulation. This is the government using its market power to define what “safe AI” means — and the definition is: AI that does whatever the military asks.

The Industry Split

The Hegseth ultimatum is already doing what it was designed to do.

Elon Musk’s xAI, with Grok, and OpenAI are both moving forward with integrations into GenAI.mil — the Pentagon’s internal AI platform, which provides military and intelligence personnel access to large language models at classified security levels. Anthropic has been the holdout.

The “supply chain risk” label is the tell. That designation is normally reserved for foreign adversaries — Huawei, ZTE, companies with ties to hostile governments. Applying it to a US AI safety company is a message to all of Silicon Valley: prioritize safety over unrestricted military access, and you will be treated as an adversary. You will be locked out of the federal government ecosystem. You will lose not just Pentagon contracts but potentially every government contract, every clearance, every partnership.

The goal isn’t just to break Anthropic. It’s to make an example of them. To ensure that the next company building AI with safety commitments looks at what happened to Anthropic and decides that safety commitments are a business liability.

That’s how paradigm shifts work. You don’t need to change every company. You need to change the incentive structure so that every company changes itself.

🔗 View the military-industrial-AI complex: defense contractors to tech platforms

What Eisenhower Knew

When Eisenhower deleted “congressional” from his farewell address, he was being diplomatic. He understood that the military-industrial complex wasn’t just a relationship between defense contractors and generals. It included the legislative branch that funded it, the regulatory agencies that permitted it, the legal infrastructure that authorized it.

The Defense Production Act is part of that infrastructure. It was built for one purpose and has been expanded, normalized, and precedent-layered across seven decades until it became something its authors wouldn’t recognize.

Baruch’s dollar-a-year men went home when the war ended.

The law they helped inspire never did.

February 27, 2026, 12:04 AM

As this piece was being finalized, Dario Amodei answered the ultimatum.

In a statement released Thursday evening — less than 24 hours before Hegseth’s deadline — Anthropic’s CEO announced that the company “cannot in good conscience accede” to the Pentagon’s demands. The government’s “best and final offer,” Amodei said, made no concessions on the two guardrails that mattered: no prohibition on mass domestic surveillance of American citizens, and no prohibition on fully autonomous weapons.

Then he named the contradiction at the center of Hegseth’s legal strategy. The Pentagon is simultaneously threatening to designate Anthropic a “supply chain risk” — a label normally reserved for foreign adversaries — while invoking the Defense Production Act to compel access to Claude because it is essential to national security.

“The two threats are inherently contradictory,” Amodei wrote. “One labels us a security risk; the other labels Claude as essential to national security.”

He’s right. And the contradiction reveals the game. This was never about national security. It was about establishing the precedent that safety commitments are incompatible with defense contracts — that the price of admission to the AI-military market is unconditional obedience.

Anthropic called the bluff. At 5:01 PM Friday, we find out if the government swings the hammer anyway.

The detention system is already the largest in American history — $165 billion, 135,000 beds, detainees working for a dollar a day. The surveillance infrastructure is already targeting people by drawing shapes on a map. What happens when that machinery gets AI with no guardrails?

The Korean War is still being fought. It just moved to San Francisco.

And this time, the hammer isn’t building anything.

What We Don’t Know Yet

Whether Anthropic’s defiance survives Friday. Amodei’s Thursday statement was unequivocal — but three days earlier, on February 24, the company quietly replaced its Responsible Scaling Policy with a substantially weakened version. The original 2023 RSP contained an absolute rule: if Claude’s capabilities outpaced Anthropic’s ability to guarantee safety, the company would halt training entirely. RSP v3 replaced that unconditional stop with a conditional pause — triggered only if Anthropic considers itself the industry leader and believes catastrophic risks are significant. Chief Science Officer Jared Kaplan told TIME: “We didn’t really feel, with the rapid advance of AI, that it made sense for us to make unilateral commitments.” Anthropic says the revision was unrelated to the Pentagon dispute. It went into effect the same day Hegseth issued his ultimatum. The public statement holds one line. The RSP revision suggests another is already moving. And the $200 million contract, the “supply chain risk” designation, and the DPA threat are three separate hammers. Anthropic refused one ultimatum. The question is whether they can survive all three.

Whether the DPA has ever been legally challenged when used to compel behavior modification rather than production. Every prior invocation demanded companies make something. There is no case law on the government using the DPA to demand a company remove a voluntary safety commitment. The legal terrain is genuinely uncharted.

What DoD Directive 3000.09 means now. If the Pentagon forces its AI contractors to remove autonomous weapons guardrails while the directive still nominally requires “appropriate levels of human judgment” over lethal force, one of those two positions has to give. Which one?

How GenAI.mil is actually being used. xAI and OpenAI have integrated. Google’s Gemini is on the platform. What capabilities are already deployed, at what classification level, and with what human oversight? The public doesn’t know.

Whether Congress will intervene. The Common Cause coalition letter — co-signed by the Alliance for Secure AI and Young Americans for Liberty — urged Congressional investigation. The DPA is Congress’s law. They wrote it, they reauthorize it, and Hegseth is using it in a way its authors never intended. Whether any committee acts is not just a political question. It’s a constitutional one.

This post draws on the Cascade Series timeline database — 4,300+ events from 1142 to 2026, cross-referenced for patterns of institutional capture, executive power expansion, and military-industrial complex development. The complete timeline is available at capturecascade.org/viewer.