The Pentagon’s Ultimatum to Anthropic Is Bigger Than One Contract

What happens when a “human-centric” AI company runs into the U.S. security state? We are about to find out.

On Tuesday, February 24, 2026, Defense Secretary Pete Hegseth sat Anthropic down and gave it a deadline. By Friday at 5:01 p.m. Eastern, the company must relax the ethical guardrails it built into its AI model, Claude, or it risks losing a Pentagon contract worth up to 200 million dollars. If Anthropic refuses, the Department of Defense is prepared to label the company a “supply chain risk,” cut it out of future work, and has openly threatened to rely on the Defense Production Act, a Cold War emergency law, to force access to its technology.

In plain English, the message is not subtle. The Pentagon wants to use Anthropic’s AI for “any lawful military use,” which explicitly includes support for autonomous weapons and large-scale surveillance. Anthropic’s policies currently say no. The Pentagon has now said change those rules or pay a price.

On the surface, this looks like a fight over one defense contract. In reality, it is a test of whether any AI company that claims to be ethical and human-centric can hold its ground when the national security apparatus decides those ethics are inconvenient. Anthropic built its identity, and wrote its terms of service, around the idea that Claude would not be used to autonomously kill people or to power dragnet surveillance of citizens. What happens in this standoff will shape not just Anthropic’s future, but how the public and policymakers decide to treat AI: as a technology that can be constrained by values, or as something that inevitably bends to power.

---

This Community Is Powered by You
What started as a small circle has grown into something much bigger, and it’s all because of readers like you.
Every time you forward this email, post it on socials, or bring someone new into the fold, you’re helping build one of the most passionate, independent political communities out there.
Want to keep the momentum going?
Share this newsletter with someone who should be part of this conversation.

Thank you for being here. It means everything.

How We Got Here

The Contract

The story begins in July 2025, when the Pentagon’s Chief Digital and Artificial Intelligence Office announced that it was awarding contracts worth up to 200 million dollars each to four frontier AI firms: Anthropic, Google, OpenAI, and Elon Musk’s xAI. The aim was to pull state-of-the-art commercial AI models into defense work at scale, from planning and logistics to reconnaissance and targeting support. Anthropic signed a two-year prototype agreement with a 200 million dollar ceiling as part of this push.

For a while, the relationship looked like a showcase for “responsible AI in defense.” Anthropic’s model Claude was the first frontier system approved to operate on classified networks. Internally, Pentagon officials praised its performance. Externally, Anthropic talked about helping the Department of Defense use AI while keeping humans in control and respecting civil liberties. It was, on paper, an experiment in marrying national security and ethics.

The Line Being Crossed

Then came Venezuela.

On January 3, 2026, U.S. forces launched a high-risk operation to capture Venezuelan president Nicolás Maduro. Reports later revealed that Claude was used during the planning and execution of the raid, integrated through a defense contractor stack. The operation ended with Maduro in U.S. custody and scores of people dead. Members of Congress, legal scholars, and human-rights observers immediately began asking hard questions about the legal basis for the raid and about how AI had been used to support it.

Anthropic, which had publicly committed to prohibiting use of its systems for autonomous weapons and mass surveillance, reportedly asked the Pentagon for clarity about what Claude had done in the Venezuela operation. That inquiry did not land well. According to multiple accounts, some Pentagon officials saw the questions as second-guessing military judgment. Tensions grew inside the Department over whether Anthropic’s safety rules were an asset or an obstacle.

The DoD Makes Demands

By mid-February, the dispute had spilled into the open. Defense officials began saying, in the background and then increasingly in public, that Anthropic’s acceptable-use policies were too restrictive for defense work. They wanted a simple standard that any lawful use of AI that military lawyers approved would be fair game, regardless of what Anthropic’s own constitution for Claude said. Then, on February 24, Hegseth made that implicit demand explicit. Loosen the restrictions, allow autonomous weapons support and broader surveillance uses, or face termination and potential compulsion under the Defense Production Act.

Anthropic updated the Claude “constitution” on January 22, 2026. However, it did not create new red lines so much as publicly restate and deepen the values the company had already promised to uphold. The Pentagon’s problem is not that Anthropic moved the goalposts in 2026. It is that those goalposts existed at all. The 2025 contract was signed under the same terms of service, with the same red line prohibitions on use.

---

Who Anthropic Says It Is

To understand why this particular ultimatum is explosive, you have to understand how Anthropic sold itself to the world.

Values/Vision

The company was founded in 2021 by former OpenAI researchers as a safety-focused AI lab. Even its name, Anthropic, signals that branding, centered on humans rather than on raw capability. From the beginning, the founders framed the company around a simple idea. Powerful AI should be developed in a way that is constrained by human values, not just by market demand or government priorities.

Structure

Anthropic did not just talk about that mission. It baked it into its legal structure. The company is organized as a public benefit corporation, which means that, on paper, its directors are obligated to consider a stated public benefit alongside shareholder profit. That is a very different message from a standard Delaware C-corp built purely around maximizing return. It is a way of telling employees, customers, and regulators that the company exists for something more than quarterly numbers.

Technology

On the technical side, Anthropic describes its approach as “Constitutional AI,” a method that trains models like Claude to follow an explicit set of high-level principles related to human rights, non-violence, and respect for privacy. The idea is that the system should internalize guardrails rather than rely only on after-the-fact filters. In Anthropic’s own marketing, Claude is not just another large language model. Instead, it is supposed to be a model that has literally been taught a kind of internal constitution.

Policies

That theme continues in the company’s public policies, which are unusually blunt for a major AI vendor. Its acceptable use rules say that Claude cannot be used to build or operate fully autonomous weapons, that it cannot be used to target individuals with violence, and that it cannot be used for bulk, suspicionless surveillance of people, especially in domestic contexts. Anthropic has said, in plain language, that it will not support applications in which AI removes humans from lethal decision loops or monitors whole populations without cause.

Their Clientele

That stance is not just moral posturing. It is part of the business strategy. Many of Anthropic’s most enthusiastic commercial users are universities, research institutes, nonprofits, and cultural institutions that want access to powerful AI tools but are wary of the worst-case scenarios such as automated killing, ubiquitous facial recognition, and mass social-media monitoring. Claude’s promise of strong guardrails and a formal public benefit mandate is, for them, a feature rather than a drawback. When they chose Anthropic over larger and more opaque contenders, it was in part because they believed those red lines, and that corporate structure, meant something.

All of that is what the Pentagon is now asking Anthropic to compromise. If Anthropic agrees, it will not just be changing a few lines in a terms of service document. It will be rewriting the core promise that made it distinct in the first place, and undermining the claim that it exists, in part, for a public benefit that includes not using its technology to help automate killing and dragnet surveillance.

---

A Cold War Law Meets a 2026 AI Lab

The other character in this story is a statute most people never think about, the Defense Production Act.

The DPA, Evolved

Congress first passed the DPA in 1950 in the early days of the Korean War. The point was to give the president tools to mobilize industrial production for national defense, specifically the power to require companies to prioritize military orders, to allocate materials, and, in some circumstances, to expand capacity for critical goods. Over the decades, Congress broadened the definition of “national defense” to include things like critical infrastructure and emergency preparedness. Administrations of both parties have used the DPA to accelerate production of everything from military hardware to personal protective equipment during COVID-19.

Those uses share a common logic. There is some clearly defined crisis or strategic need, there is not enough of a specific thing, and the government compels or prioritizes production so that need is met. Whatever you think of how far that should go, it is very much about factories, supply chains, and physical goods.

What’s Different This Time

What Hegseth has now threatened to do is very different. In this case, there is no declared war, no pandemic, and no obvious shortage of AI capacity. The Pentagon already has contracts with several AI firms, including OpenAI, Google, and xAI.. Other models are already being evaluated for use on classified networks. If Anthropic walked away tomorrow, the Department of Defense would not suddenly be devoid of AI.

The “problem,” from the Pentagon’s perspective, is not that no one can supply advanced AI. The problem is that Anthropic, one particular supplier, has ethical rules that make certain uses off limits, even if government lawyers say those uses are lawful. Invoking the DPA here would not be about ramping up scarce production in an emergency. It would be about using a 1950 emergency law to force one company to discard its own safety policies while competitors stand by with fewer objections.

Can They Do This? Yes, But…

Legally, that is murky territory. Politically, it is worse. It blurs the line between “national defense mobilization” and “punishing a company for having stronger ethics than its peers.” It sends a quiet but unmistakable signal to the entire industry. If you insist on human-centric guardrails that get in the Pentagon’s way, the U.S. government may treat your values as a national security problem.

---

Anthropic’s Fork in the Road

That brings us to the choice facing Anthropic’s leadership, which is not one any founder dreams of.

If They Are Who They Say They Are

If the company refuses to change its policies, it will almost certainly lose the Pentagon contract, along with the influence that comes from being the first frontier AI model on classified networks. It may be formally or informally blacklisted from future defense work. Its executives and board will be accused of undermining national security. There may be legal battles if the government tries to use the DPA as leverage. Investors who prize predictable government revenue will not be thrilled.

On the other hand, if Anthropic holds the line, it retains something that is very hard to buy back once lost: credibility. The universities, museums, libraries, research labs, and civil-society organizations that gravitated to Claude because of its stated limits will see that those limits are real. People who are uneasy about AI but not ready to give up on it entirely will have a concrete example of a company that chose to walk away from power rather than help automate killing and surveillance. For many, that may be the difference between “AI is inherently dystopian” and “AI can be tolerable if the right actors are willing to sacrifice for their principles.”

So far, Anthropic appears to have chosen this path.

If They Cave

If, instead, Anthropic gives in, the outcome looks very different. In the short term, it likely keeps the contract. The Pentagon gets the flexibility it wants, and the company’s technology becomes more deeply embedded in military planning and operations. Some investors will cheer.

The longer-term consequences are uglier. The institutions that chose Claude for ethical reasons will very likely start to leave. Staff who joined Anthropic because of its safety mission, especially researchers and policy professionals, will face a simple question: do they want to spend their careers building tools that support autonomous targeting and dragnet monitoring, or do they want to go somewhere else. Lawmakers and activists who already fear AI will have a powerful talking point. Even the self-described ethical lab caved as soon as the Pentagon squeezed. Calls for severe, blunt regulations and even bans on certain classes of AI will grow louder.

There is also a quieter business reality. There is only one U.S. Department of Defense, and its priorities change with elections, threat assessments, and personalities. AI contracts can be re-bid, re-scoped, or canceled. New vendors emerge. An AI firm that locks its identity and revenue model to the defense establishment may find that it has sacrificed the wider commercial ecosystem for a single, highly volatile customer.

The fork in the road looks like this. Refuse, and bet that there is a large enough community of institutions and individuals who will value a company that actually meant what it said. Cave, and bet that a future built around national security work, with shrinking public trust and growing regulation, is still worth it.

---

How Congress Let It Get This Bad

It is tempting to treat this as a purely executive branch story: the Pentagon, Anthropic, one contract, one showdown. It is not. The backdrop here is decades of congressional failure on technology in general and a very particular kind of failure on AI in the last few years.

AI Legislation (or lack thereof)

On AI, there is still no comprehensive federal statute that defines rights, responsibilities, and limits for high-risk systems. Instead, there is a scatter of older, incomplete internet laws, sector-specific rules, and a growing pile of press releases and “frameworks” that do not actually bind anyone. Into that vacuum, the political conversation in Washington has drifted toward one dominant theme: preemption of state laws in the name of “one federal standard.”

You can see that clearly in Senator Marsha Blackburn’s TRUMP AMERICA AI Act, introduced last year. On the surface, it promises a national framework for artificial intelligence. In practice, much of its energy is aimed at overriding state AI regulations that are stricter than whatever Washington ultimately agrees to. The bill reflects a familiar instinct in tech policy: treat state privacy and AI laws as “obstructions” to innovation and competition, rather than as attempts to protect residents from real harms. It gestures at oversight, but its core logic is that the federal government should set the rules and the states should largely get out of the way.

The White House has followed a similar line by executive action. In December 2025, President Trump signed an order titled “Ensuring a National Policy Framework for Artificial Intelligence.” It declares that the United States will maintain global AI leadership under a single national policy, directs agencies to challenge state laws that conflict with that goal, and calls for an AI litigation effort inside the Justice Department to go after state regulations deemed overly burdensome. The order even hints that federal funds could be used as leverage against states that refuse to fall into line.

For years, conservatives have argued that federal law should be a floor, not a ceiling, and that states should be free to go further in protecting their residents. That logic mysteriously evaporates when the subject is AI. Suddenly, the same movement that invokes the Tenth Amendment and “states’ rights” now backs federal bills and executive orders designed to keep states from regulating AI more strictly than Washington is willing to, even when Washington is unwilling to regulate at all.

The Constitution does allow federal law to preempt state law. The question is what that power is used for. You can use it to raise protections nationwide, or you can use it to stop states from protecting people where Congress has failed to act. On AI, so far, Washington has chosen the second path.

The most galling version of federal preemption is not when Washington sets a real standard and tells states they cannot tear it down. It is when Washington refuses to pass meaningful protections at all, then turns around and tells states they are not allowed to fill the gap. That is what we are drifting toward on AI: a federal “ceiling” made of memos and press conferences, paired with explicit efforts to warn states away from building their own guardrails. It is not national leadership. It is national veto power over anyone else who tries to do the job no one else was willing to do.

Mission Creep of the DPA

The Defense Production Act sits in this vacuum as a tool that Congress wrote for one era and never really revisited for this one. Lawmakers have broadened its scope over time, but without clearly spelling out what kinds of crises justify its use and what kinds do not. That ambiguity created room for exactly the kind of stretching we are seeing now, where a law passed in 1950 to prioritize steel and munitions can be floated as a way to pressure a private AI lab into dropping ethical restrictions that stand in the way of “any lawful use.”

A DoD With Little Regulation

Finally, Congress continues to fund the Department of Defense and related agencies without putting explicit, modern guardrails around AI use. There is plenty of rhetoric about keeping humans “in the loop,” but very little hard statutory language that bars federal dollars from being used to deploy systems that make or materially drive lethal decisions, or that operate as suspicionless mass surveillance tools against people in the United States. As with earlier surveillance authorities, broad mandates, vague standards, and a heavy dose of secrecy have been allowed to stand in for clear, democratically debated limits.

The conflict between Anthropic and the Pentagon is what happens when all of those failures collide with a powerful new technology. There is no clear law governing AI in war and surveillance, there is an old emergency statute that can be stretched to fit, and there are no strong conditions tied to defense funding. That leaves corporations and cabinet secretaries to hash out the boundaries in private, while Congress argues about who gets to say “states keep out” instead of doing the harder work of saying “here is what you may and may not do with this technology in the first place.”

---

What Congress Needs To Do Now

If there is any upside to this confrontation, it is that it makes the stakes impossible to ignore. Congress cannot pretend that AI is still an abstract debate while an actual company is being told to choose between its ethics and access to government money.

Real AI Regulation

First, lawmakers need to do the work they have put off for years and pass meaningful AI legislation. That does not mean a single “AI omnibus” that tries to solve everything. It does mean setting clear rules for high-risk systems, including privacy, transparency, testing, and accountability, and defining how AI can and cannot be used in sensitive domains. It also means deciding, in public, whether certain uses, like fully autonomous lethal targeting or broad domestic dragnet surveillance, should be categorically off the table.

Second, if Congress wants one national standard rather than a patchwork, it needs to write that standard in statute and make it protective enough that states do not feel compelled to go it alone. If it is not willing to do that, it should stop quietly cheering on efforts to kneecap state AI laws by executive order and litigation. Federal preemption is not supposed to mean “no one regulates this.”

Update the DPA & Connect Money to Intent

Third, Congress should modernize the Defense Production Act. It should clarify when and how the statute can be used for non-traditional domains like software and AI, and it should explicitly bar using the DPA to force private companies to abandon basic human-rights protections in their products when other suppliers exist. Emergency law is not supposed to be a shortcut for winning policy arguments that Congress has not yet had the courage to hold. It must define what constitutes an emergency.

Finally, Congress should tie money to rules. If legislators do not want AI systems quietly used against citizens in ways that stretch or ignore constitutional protections, they can say so in the next defense authorization or appropriations bill. They can condition funding on keeping humans in meaningful control of lethal decisions, on limiting AI-powered surveillance against U.S. persons, and on transparency to oversight bodies about how these tools are deployed. That is what the power of the purse is for.

---

Why This Moment Matters

It is a lot to put on one company. Anthropic did not cause the lack of AI regulation, did not write the DPA, and does not control the Pentagon’s broader adoption of AI. Yet by choosing to build a business around human-centric, constrained AI, it now finds itself at the center of a test that should have belonged to Congress years ago.

If Anthropic holds its ground and loses the contract, the company may shrink, regroup, and suffer real pain. It may also prove that it meant what it said, which would send a powerful signal to everyone watching, that ethics in AI are not necessarily doomed to collapse the moment power pushes back. If it gives in, the contract may be safe for a while, but something larger will have been lost. Those who fear AI as a tool of automated violence and surveillance will be able to say, with some justification, that even the most “ethical” lab folded when it counted.

The rest of us do not get to choose for Anthropic. We do not sit in that boardroom. What we can do is refuse to let this moment be treated as a quirky dispute between one contractor and one department. It is a stress test for the idea that advanced AI can be governed by democratic values rather than by unreviewable power.

If Congress wants that idea to survive, it has work to do. Most importantly, it must listen to the will of the people, not the interests of lobbyists.

If you want more analysis like this — grounded in names, dates, and documents, with an eye on the bigger picture rather than the daily outrage cycle — consider subscribing. We track the places where policy, power, and technology quietly redraw the lines of public life.

Sources:

• “Hegseth warns Anthropic to let the military use the company’s AI tech as it sees fit, AP sources say,” AP News, February 24, 2026.

• “Hegseth threatens to force AI firm to share tech, escalating Anthropic standoff,” The Washington Post, February 24, 2026.

• “Anthropic digs in heels in dispute with Pentagon, source says,” Reuters, February 24, 2026.

• “Hegseth threatens to blacklist Anthropic over ‘woke AI’ concerns,” NPR (via NHPR), February 24, 2026.

• “Exclusive: Pentagon threatens Anthropic punishment,” Axios, February 16, 2026.

• “Pentagon gives AI firm ultimatum: lift military limits by Friday or lose $200M deal,” Fox News, February 24, 2026.

• “Anthropic digs in heels in dispute with Pentagon, source says,” Cybernews (Reuters syndication), February 24, 2026.

• “AI research and products that put safety at the frontier,” Anthropic homepage, accessed February 24, 2026.

• “Company” (Anthropic is a Public Benefit Corporation dedicated to responsible AI for the long-term benefit of humanity) — Anthropic: Company page — updated 2025, accessed February 24, 2026.

• “Claude’s new constitution,” Anthropic News, January 22, 2026.