Anthropic Takes On The Government

On February 25, 2026, the most consequential AI alignment fight in America is not happening in a lab. It is happening across a conference table at the Pentagon, where Anthropic is refusing to loosen Claude’s safety limits even under direct pressure from Defense Secretary Pete Hegseth.

This is a story about power more than policy: who gets to decide how frontier AI is used when the customer is the U.S. government, the use cases are classified, and the stakes include surveillance and lethal force.

Hegseth has reportedly delivered an ultimatum to Anthropic CEO Dario Amodei: accept broader military access or face extraordinary government action. The Pentagon has threatened either to label Anthropic a supply-chain risk or to invoke the Defense Production Act to force compliance, with a firm deadline that turns a contract dispute into a constitutional-feeling confrontation.

Anthropic’s refusal is narrowly focused but politically explosive. The company is holding to safeguards that would prevent Claude from being used for autonomous weapons targeting and for domestic surveillance inside the United States, even as Pentagon officials argue they should only have to comply with U.S. law, not additional private-sector restrictions.

A US startup just hit a $1 billion private valuation, joining billion-dollar private companies like SpaceX, OpenAI, and ByteDance. Unlike those other unicorns, you can invest.

Over 40,000 people already have. So have industry giants like General Motors and POSCO.

Why all the interest? EnergyX’s patented tech can recover up to 3X more lithium than traditional methods. That's a big deal, as demand for lithium is expected to 5X current production levels by 2040. Today, they’re moving toward commercial production, tapping into 100,000+ acres of lithium deposits in Chile, a potential $1.1B annual revenue opportunity at projected market prices.

Right now, you can invest at this pivotal growth stage for $11/share. But only through February 26. Become an early-stage EnergyX shareholder before the deadline.

Invest in EnergyX Today

_{This is a paid advertisement for EnergyX Regulation A offering. Please read the offering circular at invest.energyx.com. Under Regulation A, a company may change its share price by up to 20% without requalifying the offering with the Securities and Exchange Commission.}

The Pentagon’s framing sounds tidy: let the military use the model for all lawful purposes, and trust internal rules and existing law to prevent abuse. But in practice, lawful is not the same as wise, rights-preserving, or democratically legitimate, especially in domains where the law never anticipated AI systems that can search, summarize, infer, and scale decision-making across millions of people and billions of records.

There is also the reliability problem. The Pentagon’s push for fewer restrictions collides with a stubborn technical reality: large language models can produce confident-sounding errors. In a military context, a plausible hallucination is not an inconvenience. It can be escalation, mission failure, or civilian harm. The argument for guardrails is not just moral, it is operational.

The dispute sharpened after reports that Claude was used during a January operation that captured Nicolás Maduro, reportedly through Anthropic’s partnership with Palantir. Whether Claude drafted summaries, accelerated analysis, or helped coordinate information flow, the central point is that once a model touches an operation like this, oversight questions become unavoidable.

What matters most is not a single mission. It is the broader institutional lesson. Anthropic’s impulse to ask how its model was used reads as what responsibility looks like in the real world: a supplier recognizing that the line between analysis support and operational enablement can disappear under urgency, secrecy, and momentum. The Pentagon’s anger, by contrast, suggests it may view any post hoc scrutiny as interference, even when the scrutiny is about preventing the most dangerous applications.

The supply-chain risk threat is not a routine contracting lever. In practice, it functions like a blacklist. If the government declares a vendor risky, contractors and partners feel pressure to purge that vendor across entire ecosystems, even where the technology is safe, legal, and widely used. A dispute over guardrails can become an attempt to isolate a company.

The Defense Production Act threat is even more alarming because it moves the fight from procurement into compulsion. It signals that if negotiation fails, the government may attempt to force access or policy changes through emergency-style authorities. Even if such a move is contested in court, the message to every AI builder is unmistakable: if your safety limits conflict with national security priorities, your limits may not survive.

The alignment community has spent years worrying that the first major test of alignment in practice would not be a neat technical failure mode. It would be a social and political stress test: powerful institutions leaning on builders to relax safeguards because safeguards are inconvenient in a race, a crisis, or a war.

LessWrong has hosted repeated discussions of government coercion and control as AI becomes a national security asset, including scenarios where national security framing drives increasing government influence over labs, and where emergency authorities and executive action reshape the playing field quickly. Some forecasting work has also explored extreme measures involving consolidation of compute and sweeping state intervention once AI is treated as strategic infrastructure. That is not the same as predicting this exact showdown by name and date, but it is the same underlying claim: once AI becomes a geopolitical asset, pressure can escalate from negotiation to coercion fast.

So no, there is not one canonical LessWrong line that declared this exact Pentagon confrontation would be the first big alignment crisis. LessWrong is a forum, not a single oracle. But the shape of the risk has been discussed for years, and the current standoff fits that shape uncomfortably well.

Anthropic is not acting like a company trying to stay pure by refusing all defense work. The argument is narrower: even in defense contexts, there should be hard limits around fully autonomous targeting and domestic surveillance.

That position is not anti-military. It is pro-democracy. It recognizes a basic institutional truth: the most durable safeguards are the ones that remain in place when incentives shift, when fear spikes, when secrecy thickens, and when the people closest to a mission are least able to say no. A system designed to spot targets faster than humans can react will also spot dissent faster than democratic checks can respond. The difference is not the model. The difference is whether we insist on constraints before capability becomes irresistible.

Anthropic’s corporate structure and public mission make this stance more than marketing. If responsible AI collapses the moment a powerful customer demands exceptions, then responsible AI was never a commitment. It was a slogan that lasted until the first crisis.

Even if you are cheering Anthropic on, a hard truth remains: a single company cannot be the Constitution. If Anthropic holds firm, the government can try to buy unconstrained systems elsewhere, and competition among vendors makes that easy.

That is why the best pro-Anthropic conclusion is not let companies decide. It is stop forcing private firms to be the only speed bump. Congress should set substantive rules for military AI, because only legislation creates durable constraints that survive vendor switching and administration changes.

If lawmakers believe mass domestic surveillance and fully autonomous weapons are red lines, they should write those red lines into law, along with transparency and reporting requirements that make trust us less central to the system. Until then, every guardrail fight will look like this: an ad hoc standoff, a deadline, and a threat.

On paper, the Pentagon can frame this as a procurement dispute. In reality, it is a referendum on whether alignment will be treated as a public safety obligation or as a removable feature once the customer has enough leverage. That is the reason the deadline feels so loaded.

Anthropic is taking the risk that saying no now matters more than any single contract. If it wins, it will prove that safety commitments can survive contact with power. If it loses, the lesson to every AI lab will be simple: the first real alignment problem was never just getting models to follow rules. It was getting the strongest institutions on Earth to tolerate rules at all.