The Pentagon Asked. Anthropic Said No. OpenAI Said Yes.

A note on sourcing: this article draws on primary statements from Anthropic and OpenAI, reporting from Fortune, TechCrunch, the Washington Post, Responsible Statecraft, and Lawfare's legal analysis, along with publicly available survey data and app store figures. Some details remain contested, and the situation is still developing. Where conclusions are drawn, they reflect a synthesis of that reading. I've done my best to represent the facts as they currently stand.

Claude generated approximately 1,000 prioritised targets in the first 24 hours of US strikes on Iran. Hours earlier, Trump had declared Anthropic a national security threat and ordered every federal agency to stop using its technology.

There is a recurring plot in American business history. A powerful institution demands unconditional loyalty from a supplier. The supplier has something the institution desperately needs. The institution decides the need is greater than the principle. The supplier disagrees.

Usually, the institution wins, the supplier folds, a quiet press release is issued, and life continues.

That is not what happened here.

In July 2025, the Pentagon awarded a group of AI companies, Google, OpenAI, xAI, and Anthropic, contracts worth up to $200 million each to supply frontier AI models for military use. The logic was straightforward: the United States was in an arms race with China for AI supremacy, and the fastest path to the front was to partner with the companies who had already run the hardest miles.

Anthropic was not just one of four. It was the only company whose model had been approved for use on classified networks, the first frontier AI company in US history to clear that bar.

Anthropic, founded in 2021 by former OpenAI researchers who left because they believed the commercialisation of AI was moving too fast relative to its safety, signed on. With conditions. Actually, two of them.

First: Claude could not be used for fully autonomous weapons systems, meaning AI making kill decisions without a human in the loop.

Second: Claude could not be used for mass domestic surveillance of American citizens.

These were not new positions. They had been Anthropic's stated red lines since before the contract was signed. The Pentagon knew. They agreed, or at least agreed to a version of the language. The contract was signed. Anthropic delivered.

Then, in January 2026, the Pentagon came back. It wanted the language removed.

Specifically, according to Amodei's own public statement, the Department of Defence asked them to delete "a specific phrase about 'analysis of bulk acquired data'": the single contractual clause that addressed the mass surveillance scenario they were most worried about. The wording of his statement is telling: "We found that very suspicious".

Anthropic said no.

Before the formal ultimatum, there was a quieter moment that set everything in motion.

On 3 January 2026, US Special Operations forces, using Claude via Palantir's Maven Smart System, the AI targeting platform embedded in classified military networks, carried out Operation Absolute Resolve. Venezuelan President Nicolás Maduro and his wife were captured in Caracas.

Shortly after, an Anthropic employee asked a Palantir counterpart how Claude had been used in the raid. An innocent question. The kind of thing someone asks when they build a model and wonder where it ends up.

The Palantir executive reported the question back to the Pentagon. The inference drawn: Anthropic's employees might not approve of how their technology was being used. The inference led to a conversation. The conversation led to a meeting. The meeting led to an ultimatum.

The whole confrontation started because someone building a language model wanted to know what it had done.

What followed was one of the more extraordinary corporate escalations in recent memory.

On 24 February, Defence Secretary Pete Hegseth met Amodei in person at the Pentagon. The tone was reportedly cordial. The message was not: sign a document giving the military full access by end of the week, or lose the contract. Three possible consequences were named, cancellation of the contract, invocation of the Defense Production Act (a Korean War-era statute that could compel Anthropic to provide Claude without restrictions), and designation as a supply chain risk to national security.

Amodei did not budge.

The Pentagon, now rebranded "The Department of War" under Hegseth, a rebrand that is not cosmetic but reflects an explicit operating philosophy, had already signalled its direction. At SpaceX's Starbase in January, Hegseth had announced an "AI-first warfighting force" and declared the Department would "not employ AI models that won't allow you to fight wars". He redefined responsible AI as "objectively truthful AI capabilities employed securely and within the laws governing the activities of the department". The word "safely" was notably absent.

The deadline was set: 5:01 p.m. ET, Friday, 27 February 2026.

On the evening of 26 February, Amodei published a lengthy public statement. Its tone was precise and uncompromising:

"We cannot in good conscience accede to their request".

"AI-driven mass surveillance presents serious, novel risks to our fundamental liberties. To the extent that such surveillance is currently legal, this is only because the law has not yet caught up with the rapidly growing capabilities of AI".

"Frontier AI systems are simply not reliable enough to power fully autonomous weapons. We will not knowingly provide a product that puts America's warfighters and civilians at risk".

And the line that cut cleanest through the political noise: "These latter two threats are inherently contradictory: one labels us a security risk; the other labels Claude as essential to national security".

5:01 p.m. passed without agreement.

Trump posted on Truth Social: "I am directing EVERY Federal Agency in the United States Government to IMMEDIATELY CEASE all use of Anthropic's technology. We don't need it, we don't want it, and will not do business with them again!" He added: "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 posted on X: "Effective immediately, no contractor, supplier, or partner that does business with the United States military may conduct any commercial activity with Anthropic".

The GSA removed Anthropic from USAi.gov, the government's centralised AI testing platform, and terminated its federal procurement schedule. Treasury, State, and HHS announced compliance. Anthropic became the first domestic company in United States history to be designated a national security supply chain risk.

That designation came with a six-month grace period, during which the Pentagon would continue using Claude. Which is where things get genuinely difficult to explain.

Within hours of Trump announcing the ban, Operation Epic Fury began. The United States and Israel launched strikes on Iran. Claude, still embedded in Palantir's Maven Smart System, was used for targeting. According to the Washington Post, Claude generated approximately 1,000 prioritised targets on the first day of operations alone, synthesising satellite imagery, signals intelligence, and surveillance feeds in real time to produce target lists complete with precise GPS coordinates, weapons recommendations, and automated legal justifications for strikes.

The government's position, stated in plain terms: we think you are a national security threat. We will keep using you to help select targets in an active war while we find someone less principled.

That someone arrived within hours.

Sam Altman had spent the day saying he agreed with Dario Amodei. On CNBC that morning, he said the Pentagon should not be threatening the Defense Production Act against AI companies, that "the few red lines that we share with Anthropic and that other companies also independently agree with" were important, and that he "mostly trusted" Anthropic as a company.

He was negotiating.

As the 5:01 deadline passed, OpenAI announced it had reached a deal.

Altman's explanation was elegant in its construction. OpenAI had the same red lines, he said. No autonomous weapons, no mass domestic surveillance. The difference, he said, was that OpenAI had agreed the Pentagon could use its technology for "any lawful purpose" and baked the limitations into the technical architecture rather than the contract language: cloud-only deployment, cleared OpenAI personnel in the loop, no edge deployment.

Whether that distinction holds depends on the appetite for legal nuance and the trust placed in Pete Hegseth's definition of "lawful".

Critics were not patient with either. Techdirt's Mike Masnick pointed out that "all lawful purposes", combined with Executive Order 12333, arguably permits NSA collection of communications data involving US persons collected outside the country. If it is technically legal, the military can use OpenAI's technology to carry it out. Altman acknowledged the complexity in an AMA two days later: "I have accepted that the US military is going to do some amount of surveillance on foreigners... I still don't like it".

Amodei did not mince words. In an internal memo that leaked almost immediately, he called OpenAI's messaging "mendacious" and "straight up lies", described Altman as "falsely presenting himself as a peacemaker and dealmaker", and called the whole arrangement "safety theatre". His characterisation of why Anthropic was sidelined went further than the contract dispute: "We haven't given dictator-style praise to Trump. We haven't donated to Trump, while OpenAI and Greg [Brockman] have donated a lot".

Altman's response, at an all-hands meeting shortly after: "So maybe you think the Iran strike was good and the Venezuela invasion was bad... You don't get to weigh in on that".

He was talking to his employees, the people building the models, and telling them that operational decisions about how their technology is used in active military campaigns are entirely up to Pete Hegseth.

OpenAI's employees were not pleased.

Chalk messages appeared outside the company's San Francisco office: "Where are your red lines?" "What are the safeguards?" "You must speak up". More than 900 employees across OpenAI and Google signed an open letter urging their employers to "refuse the Department of War's current demands for permission to use our models for domestic mass surveillance and autonomously killing people without human oversight". Multiple staff members went on record with CNN to say they "really respected" Anthropic for what it had done.

Then came the resignation Altman could not spin away.

Caitlin Kalinowski, OpenAI's head of robotics, resigned. On the record, with a statement: "Surveillance of Americans without judicial oversight and lethal autonomy without human authorization are lines that deserved more deliberation". She was specific, named, and had led one of OpenAI's most consequential teams.

Altman admitted the deal was "definitely rushed" and acknowledged "the optics don't look good". In a subsequent public AMA, he added: "If we are right and this does lead to a de-escalation between the DOW and the industry, we will look like geniuses. If not, we will continue to be characterised as rushed and uncareful".

Meanwhile: Claude shot to number one on the Apple App Store. More precisely, Claude passed ChatGPT in total daily US downloads for the first time in the company's history, hit number one in six countries simultaneously, and experienced a worldwide outage due to what Anthropic described as "unprecedented demand". Anthropic said every single day the previous week had been an all-time record for Claude sign-ups. Free active users had increased 60% since the start of the year. Daily signups had quadrupled. Paid subscribers had more than doubled.

Sensor Tower recorded a 295% surge in ChatGPT uninstalls on the Saturday after the announcement. Claude downloads were up 51% on the same day.

The consumer market had revealed a preference for AI with a spine.

OpenAI, recognising the damage, revised its contract language over the weekend to more explicitly prohibit use of commercially acquired personal data and bulk surveillance, citing the Fourth Amendment directly and explicitly barring NSA use. The autonomous weapons language remained conveniently vague. Altman said: "I got things wrong". He said the company "should not have rushed".

Amodei, now negotiating again with Pentagon CTO Emil Michael even after the formal designation, confirmed Anthropic was still providing Claude to the US military during the Iran operations, at nominal cost, "for as long as necessary" to allow a transition.

The situation as of this writing: unresolved, messy, and genuinely important.

Most coverage of this story has missed the actual point: this is not a story about government contracts. It is a story about who gets to set the conditions under which the most powerful technology in human history is deployed.

This is the Gilded Age moment for AI. When Standard Oil controlled refining, or AT&T controlled telephony, or Microsoft controlled the desktop operating system, the government was the check on corporate power. In those stories, government constrained the dominant player. In this story, the roles are reversed: the government is the customer demanding the removal of the check, and the company is the entity attempting to maintain it.

There are no villains in the classic sense. Anthropic is not naive about what it signed up for. It has had a Pentagon contract since July 2025. It worked with Palantir, who is not a civil liberties organisation. Its model was used to help plan a raid that captured a foreign head of state, and it kept providing that capability throughout the dispute. Its position was not "no military use". It was "no autonomous killing" and "no mass surveillance of your own citizens". Those are, by most reasonable standards, modest asks.

An ITIF/Morning Consult survey of nearly 2,000 American adults, published the week of the dispute, found that 79% of respondents, 81% of Democrats and 81% of Republicans, say a human should always make the final decision before any use of lethal force. Seventy per cent say AI-powered mass surveillance without a warrant violates the Fourth Amendment. Sixty-seven per cent say private tech companies have a responsibility to set limits on how their products can be used, even if the government disagrees. The public was not with the Pentagon on this. It was with Anthropic, by considerable margins, across party lines.

OpenAI is not evil for taking the contract. It is rational. The Pentagon's FY2026 budget is $1.01 trillion, larger than the GDP of most countries. Military AI is the proximate future of a significant portion of the industry's revenue. Walking away would require the kind of principled clarity that very few companies, once valued at over $700 billion, are capable of sustaining.

But the implications stretch far beyond these two companies and this one contract.

The supply chain risk designation is, legally, on thin ice.

The statute invoked, 10 U.S.C. § 3252, defines "supply chain risk" as the risk that "an adversary may sabotage, maliciously introduce unwanted function, or otherwise subvert the design, integrity, manufacturing, production, distribution, installation, operation, or maintenance of a covered system". The operative word is adversary. The statute was written for Huawei. For Kaspersky. For ZTE. For foreign entities under the direction of governments hostile to the United States. Every prior supply chain risk designation in US history has been applied to foreign companies.

Anthropic is an American company, building AI in San Francisco, whose CEO sat across a table from the Defence Secretary for two hours before declining to sign a contract amendment. That is a contractual disagreement, not adversarial sabotage.

Hegseth did not appear to follow the procedural requirements the statute demands: no joint recommendation from the Under Secretary and the Chief Information Officer, no written determination, no evidence of considering less intrusive measures, no prior congressional notification. Federal Acquisition Regulation 9.402(b) states that exclusion "shall be imposed only in the public interest for the Government's protection and not for purposes of punishment". Trump's Truth Social post calling Anthropic "radical left" and "woke" makes the purpose reasonably clear.

Courts have weighed in on similar Pentagon designations before. In Luokung Technology v. DoD (2021) and Xiaomi v. DoD (2021), federal courts granted preliminary injunctions against Pentagon supply chain designations and found them arbitrary and capricious. Those companies at least fit the statutory definition; they were foreign entities. Anthropic does not.

Lawfare published a detailed legal analysis concluding the designation "won't survive first contact with the legal system". Dean Ball, former Trump White House AI adviser and one of the architects of Trump's AI action plan, not a liberal critic, called it "attempted corporate murder" and "almost surely illegal". He said it sent "a terrible message to any business about whether it should ever risk doing business with the United States government".

Hegseth's broader order, that "no contractor, supplier, or partner that does business with the United States military may conduct any commercial activity with Anthropic", almost certainly exceeds the statute's scope. The law was written for covered procurement actions, not to regulate contractors' private commercial relationships. Microsoft's legal team reviewed the designation and concluded it could continue working with Anthropic on non-defence projects. Amazon and Google made similar assessments for their enterprise customers.

And yet.

It will take years to resolve in court. In the meantime, every general counsel at every Fortune 500 company with any Pentagon exposure will ask one question: is using Claude worth the complication? Lockheed Martin has already said it will seek alternative AI providers. The GSA has removed Anthropic from federal procurement schedules. Several Fortune 500 companies reportedly convened emergency meetings to audit their Claude dependencies.

That is the actual weapon: not the designation itself, but the chilling effect, the corporate hesitation, the quiet migration to OpenAI by procurement departments who want fewer complicated conversations. This is how governments have always operated when they cannot legally break a company: they make the company too complicated to partner with.

The question nobody wants to ask out loud: if this worked, what does the next company hold the line on? And what happens to that company after that?

Dario Amodei is the most interesting protagonist in American corporate life right now.

He left OpenAI in 2021 because he believed it was moving too fast. He founded a company built around Constitutional AI, literally writing down the principles by which the model is trained to behave, including the instruction that Claude should refuse to assist actions that "would help concentrate power in illegitimate ways. This is true even if the request comes from Anthropic itself". He has published extensively on the civilisational risks of AI developed without guardrails. He is, by all accounts, genuinely committed to the position he is taking.

He is also the CEO of a company that has raised over $12 billion, is valued at approximately $380 billion as of February 2026, generates $14 billion in annual revenue, and counts Amazon ($8 billion invested) and Google among its major backers. Those investors have their own government contracts, their own political relationships, and their own tolerance for the kind of risk that comes with taking the US military to court.

There are two ironies that complicate the portrait.

The first: the same week as the Pentagon standoff, Anthropic updated its Responsible Scaling Policy, removing the binding commitment to pause training if safety measures could not keep up with capabilities. Critics noted the timing. The company holding the line on autonomous weapons and mass surveillance quietly softened its internal safety commitments in the same news cycle.

The second: Amodei apologised for tone, not substance. The leaked memo calling Altman a liar, calling the deal "mendacious", accusing OpenAI of "straight up lies", within six days Amodei described it as written "within a few hours" of cascading announcements and an "out-of-date assessment". He was, apparently, also negotiating. Even as the formal supply chain risk designation arrived in writing on 5 March, Amodei confirmed he was in "productive" ongoing conversations with Emil Michael about reopening talks.

The question is not whether Amodei believes what he is saying. He almost certainly does. The question is how long a principled stand can be sustained when the entity standing against it can designate a company a national security threat, instruct every defence contractor to stop doing business with it, and give its biggest competitor the contract it just lost, all before lunchtime on a Friday.

The answer, historically, is: not very long.

OpenAI deserves a longer examination than it is getting.

The company was founded on principles about the safe development of artificial general intelligence. It converted from a non-profit to a capped-profit structure in 2019, raised billions, and launched ChatGPT. It completed its full commercial conversion to a Public Benefit Corporation in October 2025, at which point it removed the word "safely" from its mission statement, the word that had appeared in every IRS filing the company had ever made. The nonprofit foundation now holds approximately 26% of the for-profit entity, down from majority control. The board no longer has authority to revoke investor equity based on safety concerns.

That was not the only quiet removal. In 2024, before the Pentagon contracts were signed, OpenAI removed language from its usage policies barring "activity that has high risk of physical harm, including weapons development and military and warfare". The change was made without announcement.

The through-line of OpenAI's journey is a recurring pattern: a statement of principle, followed by a business need that tests the principle, followed by a reframing of the principle that accommodates the business need. The technology becomes essential, the company becomes powerful, and the principles become negotiable.

The Pentagon contract is the clearest example yet. Altman said he had the same red lines as Amodei, took the contract on different terms, then told his employees they do not get to have opinions about whether those terms hold in practice. He admitted it was rushed, his robotics lead resigned, and he revised the contract over the weekend.

He may be right that OpenAI's technical architecture provides better protections than Anthropic's contractual language. Cloud-only deployment, cleared personnel in the loop, the Fourth Amendment cited directly; these are genuine arguments.

But the credibility gap is real. And in AI, credibility is the only currency that ultimately matters. The systems being built now will make decisions, medical, legal, financial, military, that were previously made by humans accountable to other humans. The case for trusting those systems rests entirely on the demonstrated trustworthiness of the people building them. The case collapses the moment the people constructing those systems cannot be trusted to mean what they say about what those systems will and will not do.

Ilya Sutskever, OpenAI's co-founder who left the company in 2024, posted on X on the evening of 27 February, before the OpenAI deal was public: "It's extremely good that Anthropic has not backed down, and it's significant that OpenAI has taken a similar stance". He was wrong about the second part. He may have known it when he posted.

The Pentagon will get what it wants, or close to it. Not because it is right, but because it controls too much of the financial oxygen.

Anthropic will settle, or lose in court, or find a form of words that allows both parties to declare a partial victory. The supply chain risk designation will eventually be lifted, quietly, after the news cycle has moved on. Amodei is already talking to Emil Michael, the Iran operations continue, and the model keeps running.

OpenAI will sign more government contracts. The language will get cleaner over time, because public pressure and staff revolt are real inputs. But the precedent, that an AI company can be blacklisted for maintaining ethical limits, has been set. The next company that tries will do so knowing what happened to the last one.

The more interesting question is what happens next in the market.

Claude was number one on the App Store the week it was designated a national security threat. ChatGPT uninstalls surged nearly 300% in a single day. Anthropic's daily signups quadrupled. That is a data point worth examining carefully. The consumer market revealed a strong preference for AI with a demonstrated spine. That preference has commercial consequences. The company that held the line is growing faster than it ever has.

The industry is entering the phase that every powerful technology sector enters eventually: the phase where the money gets large enough that the principles get complicated. Standard Oil, Bell Telephone, Microsoft, Google. The pattern is consistent.

In those earlier cases, the government was the check. In this one, it is the pressure. Which changes the calculus considerably.

The only variable is how long the negotiation takes.

For most of us watching from the southern hemisphere, this looks like someone else's problem. It is not.

New Zealand is, in the words of US officials, "an integral part of the alliance's kill chain". NZ intelligence already enables armed drones to locate human targets. Under AUKUS Pillar Two, which covers AI, quantum computing, and autonomous systems, New Zealand faces increasing pressure to align its defence technology posture with Washington's. The Five Eyes alliance, of which NZ is a founding member, presupposes interoperability on precisely the kinds of systems now at the centre of this dispute.

New Zealand businesses that use Claude through AWS, or through any platform with US defence contractor exposure, are collateral in this dispute. The Mayer Brown legal analysis advises contractors to "audit Anthropic exposure immediately" and prepare justifications for continued use. Microsoft and Amazon have both concluded that the designation, as currently written, applies only to direct Pentagon contract work, not to commercial use. But that legal conclusion has not been tested in court, and "has not been tested in court" is not a phrase that general counsel like to rely on.

Every company building on Anthropic's API should be watching the legal outcome. If the supply chain risk designation holds, or if the commercial activity restriction is upheld beyond its likely statutory scope, the downstream implications for how AI tools are contracted and deployed globally will be significant.

More importantly: the companies building the AI we all rely on are now engaged in a public, contested negotiation about what that AI will and will not do when governments ask. The negotiation is happening in real time, with real consequences, using real military operations as the test environment.

The outcome will set terms that outlast any single administration, any single contract, and any single company's willingness to hold a line.

Track the legal challenge. Audit Anthropic dependencies across the business. Watch what Amodei does next.

Article current as of 8 March 2026. The legal challenge is ongoing. Negotiations between Anthropic and the Pentagon are ongoing. Claude is still running in Palantir's Maven Smart System. The Iran campaign is ongoing.

Written by Mike Bayly

Primary statements

News reporting

App store and download data

Legal analysis

Public opinion

OpenAI mission and policy changes