#Pentagon, Anthropic, and the Battle Over Who Gets to Build Military AI

On February 27, 2026, Secretary of Defense Pete Hegseth designated Anthropic a "supply chain risk to national security" — the first time in U.S. history that designation has been applied to an American company. The trigger was Anthropic's refusal to remove two safety guardrails from Claude: a prohibition on fully autonomous weapons systems and a prohibition on mass domestic surveillance. Anthropic sued the Trump administration on March 9, calling the designation "unprecedented and unlawful." Microsoft filed an amicus brief on March 10 urging a federal judge to block the Pentagon's designation, warning that the ban would immediately force it and other contractors to dismantle existing military AI configurations. A separate brief from 37 researchers at OpenAI and Google backed Anthropic the same day. Palantir confirmed it is still using Claude in active military operations in Iran — even while the blacklist plays out in court. OpenAI, meanwhile, quietly signed a competing Pentagon deal. This post covers the full timeline, the legal arguments, the companies that sided with Anthropic, and what it means for the future of military AI.

#How the Most Consequential AI Policy Fight of 2026 Began

The story officially starts in January 2026, when Defense Secretary Pete Hegseth issued an AI strategy memorandum directing all Department of Defense AI contracts to incorporate standard "any lawful purpose" language within 180 days. The instruction contradicted the contract Anthropic had already signed with the Pentagon — a transaction agreement with a $200 million ceiling, awarded in July 2025, that included specific usage restrictions.

But the friction actually started earlier.

A few months prior, Anthropic had learned through a conversation with Palantir — the data analytics firm that serves as the bridge between Claude and the Pentagon's classified networks — that Claude had been used in a U.S. military raid that displaced Venezuelan President Nicolás Maduro. The raid involved "kinetic fire." People were shot.

Anthropic's team raised the issue with Palantir in a way that a senior Pentagon official later described as implying the company "might disapprove" of their technology being used in an operation where lethal force occurred. The Pentagon read that inquiry as a warning sign. "Beyond the Maduro incident," the official told Axios, there was a broader "culture clash" — describing Anthropic as the "most ideological" of the AI labs on the question of AI risk.

Anthropic's response: the company flatly denied ever discussing the use of Claude for specific operations with the Department of War.

The stage was set for a confrontation neither side appeared willing to avoid.

#The Deadline, the Refusal, and the Blacklisting

#February 24–27: Four Days That Changed AI Policy

On February 24, Hegseth gave Anthropic CEO Dario Amodei a deadline — relent by 5:01 p.m. on Friday, February 27. Remove the restrictions that barred Claude from being used for mass domestic surveillance of Americans and for fully autonomous weapons systems. Agree to the military's standard "any lawful use" language. Or face consequences.

On February 26, Anthropic published its response. Amodei wrote that the company "cannot in good conscience" grant the DoD's request. He was specific about why.

On autonomous weapons: current AI systems "are simply not reliable enough to power fully autonomous weapons," Amodei wrote. They "cannot be relied upon to exercise the critical judgment that our highly trained, professional troops exhibit every day." On surveillance: AI systems risk piecing together "scattered, individually innocuous data into a comprehensive picture of any person's life" — a risk that Anthropic was not willing to bless contractually.

Amodei also pointed out the logical contradiction in the government's position: "These latter two threats are inherently contradictory — one labels us a security risk; the other labels Claude as essential to national security."

The Friday deadline arrived. Anthropic did not budge.

President Trump directed every U.S. federal agency to "immediately cease" all use of Anthropic's technology, adding on social media: "Anthropic better get their act together, and be helpful during this phase out period, or I will use the Full Power of the Presidency to make them comply, with major civil and criminal consequences to follow." Hegseth declared Anthropic a supply chain risk — a designation historically reserved for foreign adversaries like Huawei.

It was the first time in U.S. history the label had been applied to an American company.

#The Pentagon's Case, and Why Legal Experts Are Skeptical

The Pentagon's official position is that Anthropic's guardrails are unnecessary because the uses Anthropic is concerned about are already prohibited by federal law and existing DoD policy. Pentagon CTO Emil Michael told CBS News that the military had offered to put in writing that it specifically acknowledges federal surveillance laws and existing autonomous weapons policies — but said the DoD would "never say in writing to a company" that it won't be able to defend itself in the future.

"At some level, you have to trust your military to do the right thing," Michael said. He also pointed to China's rapidly advancing military AI program as the reason the Pentagon cannot afford to operate with restrictions it must justify to contractors.

Anthropic's counterargument was blunt. New contract language it received overnight from the Pentagon "made virtually no progress on preventing Claude's use for mass surveillance of Americans or in fully autonomous weapons." The language, Anthropic's spokesperson said, would allow its guardrails to be "disregarded at will."

Legal experts have sided largely with Anthropic on the legal merits. Reuters reported that lawyers suggest the government's case may be undermined by three factors: a mismatch between the law invoked and Anthropic's actual conduct, internal contradictions in the Pentagon's behavior, and evidence that the designation may have been driven by animus over the Maduro inquiry rather than genuine security concerns.

Alan Z. Rozenshtein of the University of Minnesota Law School wrote in Lawfare that invoking a Cold War-era production statute to force an AI company to remove safety constraints it voluntarily built "is too important a question to be resolved" that way. "The question of what values to embed in military AI" requires Congress, not a contract dispute.

#The Lawsuit, the Amicus Briefs, and the Industry Response

#Anthropic Sues — March 9

On March 9, Anthropic filed suit against the Trump administration in U.S. District Court in San Francisco, calling the Pentagon's designation "unprecedented and unlawful" and warning of "hundreds of millions of dollars" in jeopardized contracts. The company described the action as retaliation for its principled refusal — not a genuine national security determination.

#Microsoft Files — March 10

On March 10, Microsoft filed an amicus brief urging the federal judge to issue a temporary restraining order blocking the Pentagon's designation for all existing contracts. The move was significant for two reasons.

First, Microsoft is an Anthropic competitor and a company with its own substantial Pentagon contracts. Filing on Anthropic's side means accepting some political risk with the current administration. Second, Microsoft's brief revealed the operational stakes in concrete terms: without a restraining order, Microsoft and other government contractors would need to "act immediately to alter existing product and contract configurations" used by the Defense Department.

"AI should not be used to conduct domestic mass surveillance or put the country in a position where autonomous machines could independently start a war," Microsoft said in its filing.

Former Trump AI policy adviser Dean Ball called Microsoft's intervention surprising: "I would not have predicted this and I think Microsoft deserves serious praise here."

#37 OpenAI and Google Researchers File Separately

A separate amicus brief, filed the same day by 37 researchers and engineers from OpenAI and Google, added unusual weight to Anthropic's position. These are employees of Anthropic's direct competitors, filing in support of the company in a federal case against the U.S. government. The convergence signals that the broader AI research community views the Pentagon's designation as setting a precedent that could affect any company that declines military contracts on safety grounds.

#The OpenAI Factor: A Different Deal, a Different Line

The most uncomfortable subplot in this story involves OpenAI. On February 27, 2026 — the same day Anthropic was blacklisted — OpenAI signed its own Pentagon contract to deploy its AI systems in classified military environments.

Sam Altman publicly backed Anthropic's stance on autonomous weapons and mass surveillance. He wrote that the Pentagon "agrees with these principles, reflects them in law and policy, and we put them into our agreement." OpenAI's contract includes three explicit red lines: no use for mass surveillance of Americans, no directing of autonomous weapons where policy requires human control, and no use in "high-stakes autonomous decisions" that require human approval.

The critical difference between OpenAI and Anthropic is not, on paper, about the guardrails themselves. Both companies publicly oppose autonomous weapons and mass surveillance. The difference is contractual language. Anthropic sought outright bans on those specific uses. OpenAI accepted "any lawful purpose" framing with its red lines baked into the contract as stated commitments rather than hard prohibitions.

Altman closed OpenAI's statement with a notable ask: "We are asking the DoW to offer these same terms to all AI companies, which in our opinion we think everyone should be willing to accept. We have expressed our strong desire to see things de-escalate."

#What's Actually Happening Right Now: The Live Contradiction

The strangest part of the entire situation is this: Claude is still being used by the U.S. military in active operations.

Palantir CEO Alex Karp confirmed on March 12 that Palantir is still using Anthropic's Claude as the Pentagon blacklist plays out in court. Claude was the first model the Pentagon brought into its classified networks, and a senior defense official confirmed to CBS News that it was being used in military operations in the war in Iran — even after the blacklist was issued.

The Pentagon's own memo, issued March 6 and signed by Chief Information Officer Kirsten Davies, ordered all military commanders to remove Anthropic AI from key systems — including those for nuclear weapons, ballistic missile defense, and cyber warfare — within 180 days. But 180 days is not immediately. And the court restraining order Anthropic and Microsoft are seeking would pause even that clock.

The Pentagon memo warned that adversaries "can exploit vulnerabilities" in daily operations, and that possible exploitation could pose "potential catastrophic risks to the warfighter" — while simultaneously granting a six-month wind-down period and acknowledging that Claude was embedded in mission-critical systems.

The contradiction the Pentagon is now living with is the one Amodei identified in his February 26 statement: labeling Claude a security risk while also acknowledging it is essential to national security operations — and can't be quickly replaced because "the other model companies are just behind."

#Why This Case Will Define the Future of Military AI

The Anthropic-Pentagon dispute is not, at its core, a contract dispute. It's a constitutional and ethical question about who has the authority to set limits on how AI can be used in warfare and domestic intelligence — and whether companies that build the technology have standing to draw those lines at all.

Anthropic's argument is that a private company building frontier AI systems has not just a right but an obligation to embed values into those systems — values that reflect democratic principles and international humanitarian law. Amodei has said explicitly: "I believe that we have to defend our country from autocratic adversaries like China and Russia. But as we defend ourselves, we have to do so in ways that defend and preserve our democratic values."

The Pentagon's argument is that in a free society, elected governments — not private companies — have democratic authority over the conduct of war and national security operations. "At some level, you have to trust your military to do the right thing," Michael said.

Both arguments have democratic legitimacy. The question of which one should prevail in a federal court — and in the court of public opinion — is something that has never been adjudicated before. The Anthropic case is the first.

#What This Means Right Now

For AI companies: The designation makes clear that any AI company working with the Pentagon that refuses a government demand on safety grounds can face this kind of retaliation. The Microsoft and OpenAI researcher briefs suggest the industry is trying to close ranks against that precedent — but OpenAI's own deal with the Pentagon shows the line between solidarity and self-interest is real.

For enterprise customers of Anthropic: Existing contracts for non-defense work are unaffected. Microsoft, Amazon, and Google all confirmed that access to Anthropic's models through their cloud platforms remains operational. The 180-day wind-down applies only to Pentagon and defense contractor work.

For anyone who uses Claude: The question of whether AI companies should be allowed to build hard limits into their products — limits that cannot be overridden by any customer, including governments — is now being decided in a federal courtroom. That outcome will shape what every AI product looks like in every domain, not just military AI.

For policymakers watching: The Senate Armed Services Committee sent a private letter to both Anthropic and the Pentagon urging resolution before the Friday deadline. It didn't work. Lawmakers on both sides of the aisle have since called for Congress to establish clear statutory rules on military AI use. The court case may force their hand.

#FAQ

What are the two guardrails Anthropic refused to remove?
Anthropic's two non-negotiable positions are: Claude must not be used for mass domestic surveillance of American citizens, and Claude must not be used to power fully autonomous weapons systems — meaning AI makes lethal targeting decisions without a human in the decision loop.

What is a supply chain risk designation?
It is a national security label that, when applied to a company, requires all defense contractors, vendors, and government agencies to certify they are not using that company's products in their Pentagon-related work. It has historically been applied to foreign companies — most famously Huawei. Applying it to an American company is unprecedented.

Is Claude still being used by the U.S. military?
Yes, as of March 12, 2026. Palantir CEO Alex Karp confirmed Palantir is still using Claude in active operations while the court case is pending. Claude was embedded in classified Pentagon systems and is being used in military operations in the war with Iran. The Pentagon's own memo gives a 180-day removal window.

Why did OpenAI sign a Pentagon deal when Anthropic was being blacklisted?
OpenAI's position is that its deal includes contractual language that upholds the same values Anthropic held firm on — no mass surveillance, no fully autonomous weapons — but framed as stated commitments within an "any lawful purpose" contract structure rather than as outright prohibitions. The distinction is largely about how the guardrails are enforced. Anthropic wanted the limits written as restrictions. OpenAI accepted them as policy statements.

What is Microsoft's stake in this case?
Microsoft integrates Anthropic's products — particularly Claude — into technology it provides to the U.S. military. Without a temporary restraining order, Microsoft would have been forced to immediately alter existing contract configurations with defense customers. It also serves as a major distribution partner for Anthropic through Azure's cloud platform.

What happens next in court?
The federal court in San Francisco is considering Anthropic's motion for a temporary restraining order. If granted, it blocks the Pentagon designation while the case proceeds. If denied, Anthropic and all contractors using Claude in defense work must begin a 180-day transition. The outcome will set a legal precedent for whether the government can designate an American AI company a national security risk as retaliation for refusing to remove safety features.