In its fight with Hegseth, Anthropic confronts perhaps the biggest crisis in its five-year existence | Fortune

Pete Hegseth, the U.S. secretary of warfare, has demanded that Anthropic take away restrictions it presently stipulates in its contracts that prohibit its AI fashions getting used for mass surveillance or from being included into deadly autonomous weapons, which may make selections to assault with out human intervention. As an alternative, Hegseth needs Anthropic to stipulate that its expertise can be utilized for “any lawful goal” that the Division of Struggle needs to pursue.

If the firm doesn’t comply by Friday, Hegseth has threatened to not solely cancel Anthropic’s current $200 million contract with his division, however to have the firm labelled a “provide chain danger,” that means that no firm doing enterprise with the Division of Struggle could be allowed to make use of Anthropic’s fashions. That would eviscerate Anthropic’s progress—simply as the firm, which is presently valued at $380 billion, has been seeing vital business traction and is considering an preliminary public providing as quickly as subsequent yr.

A Tuesday assembly between Hegseth and Anthropic CEO Dario Amodei in Washington, D.C., didn’t resolve the battle and ended with Hegseth reiterating his ultimatum.

The dispute comes towards a backdrop of typically overt hostility in the direction of Anthropic from different Trump administration officers. AI czar David Sacks in explicit has publicly attacked the firm on social media for representing “woke AI” and the “doomer industrial advanced.” Sacks has accused the firm of participating in a “subtle regulatory seize technique primarily based on fearmongering.” His argument is principally that Anthropic executives disingenuously warn of maximum dangers from AI programs in order to justify rules on the expertise with which solely Anthropic and some different AI corporations can simply comply.

Anthropic CEO Dario Amodei has known as such views “inaccurate” and insisted that the firm shares many coverage targets with the Trump administration, together with desirous to see the U.S. stay at the forefront of the improvement of AI expertise.

Nonetheless, Sacks and others inside the administration could also be hoping Hegseth makes good on his threats to blacklist Anthropic from the nationwide safety provide chain.

Different AI corporations, similar to OpenAI and Google, have apparently not imposed restrictions on how the U.S. army makes use of their tech.

Rules versus pragmatism

Working with the army has been controversial amongst some expertise staff. In 2018, Google confronted a vocal workers insurrection over its determination to assist the Pentagon with “Venture Maven,” an effort to make use of AI to investigate aerial surveillance imagery. The worker revolt compelled Google to drag out of a bid to resume its contract to work on the undertaking. However in the years since, the web big has quietly renewed its ties with the protection institution, and in December, the Division of Struggle introduced it could deploy Google’s Gemini AI fashions for a variety of use instances.

Owen Daniels, affiliate director of research at the Heart for Safety and Rising Expertise (CSET) at Georgetown College, advised the Related Press that “Anthropic’s friends, together with Meta, Google and xAI, have been prepared to conform with the division’s coverage on utilizing fashions for all lawful purposes. So the firm’s bargaining energy right here is proscribed, and it dangers dropping affect in the division’s push to undertake AI.”

However rules could also be an unusually highly effective motivator for Anthropic workers. The corporate was based by a bunch of researchers who broke away from OpenAI in half as a result of they have been involved that lab was permitting business pressures to divert it from its unique mission of guaranteeing highly effective AI is developed for humanity’s profit. And extra not too long ago, Anthropic staked out principled positions on not incorporating promoting into its Claude merchandise and never creating chatbots particularly designed to be romantic or erotic companions.

Given the firm’s tradition, some outdoors commentators have speculated that at the very least some Anthropic workers will resign if the firm offers in to Hegseth’s calls for and drops the limitations presently constructed into its authorities contracts.

Hegseth has additionally stated there may be another choice obtainable to the Pentagon if Anthropic doesn’t comply with its request voluntarily. This could contain utilizing the Protection Manufacturing Act of 1950 to compel Anthropic to supply the army a model of its Claude mannequin with none restrictions in place. 

That DPA, which was initially designed to permit the authorities to take cost of civilian manufacturing in the occasion of warfare, was invoked throughout the Covid-19 pandemic to compel corporations to supply protecting gear and vaccines. Since then, it has been used quite a few occasions, largely by the Biden administration, even in the absence of a transparent nationwide emergency. As an illustration, in 2023 the Biden White Home invoked the DPA to drive tech corporations to share details about the security testing of their superior AI fashions with the authorities.

Katie Sweeten, who served till September 2025 as the Division of Justice’s liaison to the Division of Protection and is now a companion at the regulation agency Scale, advised CNN that Hegseth’s place didn’t make sense from a coverage perspective. “I might assume we don’t need to make the most of the expertise that’s the provide chain danger, proper? So I don’t understand how you sq. that,” she stated.

Dean Ball, who served as an AI coverage advisor to the Trump Administration, serving to to draft its AI Motion plan, and who’s now a senior fellow at the Basis for American Innovation, additionally known as the Pentagon’s place “incoherent” in a submit on X. “How can one coverage possibility be ‘provide chain danger’ (normally used on international adversaries) and the different be DPA (emergency commandeering of essential property)?” he stated.

Ball advised Tech Crunch that imposing the provide chain danger label would ship a horrible message to any firm doing enterprise with the authorities. “It might principally be the authorities saying, ‘When you disagree with us politically, we’re going to attempt to put you out of enterprise,’” he stated. 

Some authorized commentators famous that either side of the dispute had some reliable arguments. “We wouldn’t need Lockheed Martin promoting the army an F-35 after which telling the Pentagon which missions it might fly,” Alan Rozenshtein, an affiliate professor of regulation at the College of Minnesota and a fellow at Brookings, stated in a column posted on the website Lawfare.

However Rozenshtein additionally argued that Congress, not the Pentagon, ought to set the guidelines for a way the U.S. army deploys AI. “The phrases governing how the army makes use of the most transformative expertise of the century are being set by means of bilateral haggling between a protection secretary and a startup CEO, with no democratic enter and no sturdy constraints,” he wrote.

As of midweek, Anthropic confirmed no indicators of backing down from its place.

Claude’s future at stake

Helen Toner, the interim government director of Georgetown’s CSET and a former OpenAI board member, posted on X that the Pentagon is probably going underestimating the extent to which Anthropic could also be reluctant to desert its place as a result of—as bizarre as this sounds—doing so would possibly set a nasty instance for future variations of Claude. Anthropic researchers have more and more voiced issues about what every successive model of Claude learns about its personal character primarily based on coaching knowledge that now consists of information articles and social media commentary about Claude itself. 

However the firm has compromised earlier than when its again has been towards the wall. In June 2025, Anthropic confronted a doubtlessly existential menace when a federal choose dominated that its use of libraries of pirated books to coach its Claude AI fashions was probably a violation of copyright regulation. This left the firm dealing with tens of billions of {dollars} in potential liabilities if it took the case to a full trial and misplaced. As an alternative of continuous to fight the case, Anthropic introduced a $1.5 billion settlement with the copyright holders.

And simply this previous week, Anthropic demonstrated once more, in a special context, that it’s typically prepared to place pragmatism and business imperatives forward of high-minded rules. The corporate up to date its Accountable Scaling Coverage (RSP), dropping a earlier dedication to by no means practice an AI mannequin until it might assure it had sufficient security controls in place. The brand new RSP as an alternative merely commits Anthropic to matching or surpassing the security efforts being made by opponents. It additionally says Anthropic will delay creating fashions if the firm believes it has a transparent lead over the competitors and it additionally thinks the mannequin is coaching presents a major catastrophic danger. Jared Kaplan, Anthropic’s head of analysis, advised Time that “unilateral commitments” not made sense if “opponents are blazing forward.”

Whether or not Anthropic will make an analogous concession to business pressures in its fight with the Division of Struggle stays to be seen.