Publicly, the Pentagon refused to budge. In interviews, Pentagon officials suggested that the animosity between the department and Anthropic began when the Silicon Valley company began making what the military considered ridiculous demands about how the Claude for Government model, a special version of a popular public artificial intelligence model, should be used. Senior officials said Anthropic submitted 25 pages of restrictions to them.
After months of wrangling, the 25 pages were whittled down to two restrictions. At a Pentagon meeting in February, Amodei told Hegseth that the company would not provide any products to the military unless the military gave clear assurances that the products would not be used for domestic surveillance of Americans or to produce fully autonomous weapons that would ultimately operate without human oversight.
But that didn't solve the problem. The defense secretary took a tough stance, pointing out that Raytheon has no right to decide how its missiles are used against enemies; Hegseth said those decisions were entirely within the purview of government officials, not corporate executives. When Amodei refused to budge, the Pentagon declared the company a "supply chain risk" — and banned military use.
In May of this year, Emile Michael, the undersecretary of defense for research and engineering, told a New York Times reporter that Anthropic brought it upon itself and that the Pentagon’s overreliance on a single supplier of artificial intelligence programs was a mistake, even if its products appear to be ahead of competitors and China. Asked what the company could do to get back into the Defense Department's good graces, Michael said: "Just don't expect it at the War Department."
Pentagon officials say they are nearly done removing Anthropic from military computing systems. According to officials, nearly 100 percent of the military systems that used "Claude for Government" have replaced Anthropic's products with other leading-edge models - versions that have been modified for use in classified networks. In addition, the Pentagon has stopped using all Anthropic tools in its Maven system - which is used to analyze intelligence and provide target recommendations for air strikes in the war against Iran.
Military contractors affected by the ban have been reluctant to talk about the matter, perhaps fearing offending their largest customers. In a statement, a Lockheed Martin spokesman declined to comment directly on the letter it received requesting the removal of Anthropic, but said "we are following the direction of the President and the War Department." The spokesperson added that it expects "minimal impact" to its business because none of its work relies on a single large language model, without naming Anthropic.
But the Pentagon's break with Anthropic is not complete.
Just two months after Amodei's falling out with Hegseth, Anthropic launched Mythos, an artificial intelligence product uniquely suited to finding and exploiting software vulnerabilities. The Trump administration briefly banned all foreigners from accessing the program — now looking like an overreaction — and then reversed the ban.
That puts the Pentagon in a bind: It just banned itself from using what may be the most powerful new cyber weapon in history.
“Panic set in,” said one former official who recently left the government.
While the Pentagon does not allow Claude to be used for intelligence analysis or code writing, officials said the NSA has been allowed to use certain versions of Anthropic's advanced Mythos model in an "experimental" manner. The agency is both testing U.S. military computer network defense capabilities and looking for potential weaknesses in the networks of adversaries ranging from China to Russia, North Korea and Iran.
Rethinking opposition to AI regulation
This inconsistency is not limited to the Pentagon; similar situations abound at the Commerce Department, Treasury Department and the White House. As Trump's aides come and go and departments scramble for a greater voice, the administration's decisions on everything from export controls to national security reviews of artificial intelligence models are driven more by impulse and internal influence than a clear set of principles.
During its first year in office, the government's approach has been hands-off, following the advocacy of venture capitalist David Sacks - who was responsible for artificial intelligence and cryptocurrency policy before leaving the government earlier this year.
After releasing an AI action plan last year, the Trump administration slashed regulations and encouraged exports, hoping to make the world dependent on what Sacks calls "the American technology stack." In May this year, after being persuaded by Sacks and others, Trump withdrew a draft executive order requiring artificial intelligence companies to submit models to the government for security review, believing that the 90-day review requirement was too strict and would give Chinese competitors an advantage.
But officials and others say Mythos' power unnerves those within the government. By June of this year, the government implemented the toughest controls ever in the AI industry, cutting off access to the Anthropic model to foreigners, including some of the engineers who developed it. Two weeks later, it lifted the ban on foreigners without a public explanation.
There's been a similar back-and-forth in dealing with Nvidia chips that are critical to developing new models and building data centers. In April last year, the Trump administration halted the export of an NVIDIA artificial intelligence chip to China. It was one of the last batch of artificial intelligence chips previously approved for sale in China.
But Trump changed his stance last July after meeting with Nvidia CEO Jensen Huang in the Oval Office. Trump said the sales, previously seen as a national security threat, could be allowed as long as the U.S. government got a piece of the action.
Likewise, the administration has reversed a rule put in place by the Biden administration that ensured the most powerful chips could be sold freely only to close U.S. allies, with restrictions on sales to other countries. Trump administration officials have said there will be alternatives, but they have been at loggerheads over alternatives. Meanwhile, some U.S. companies have concluded that now that the rules have been lifted, they are free to sell advanced chips to some Chinese companies operating outside China.
Trump later visited China to meet with Xi Jinping, and Huang and other industry executives also traveled to China.
The administration is currently grappling with conflicting security considerations and political winds. If it controls emerging artificial intelligence pioneer labs too tightly and threatens to cut off foreign access, it could push Europeans toward China's model. Failure to regulate it could be a recipe for disaster, especially if more incidents occur that raise public concerns that the technology is getting out of control.
Chris McGuire, a senior fellow at the Council on Foreign Relations and a former Biden administration official, said the Trump administration's posture on export controls has been "inconsistent."
He said the government is now concerned about competition from China's open-source models, but has previously rolled back many measures that had hindered China's AI development. And it directed its most severe restrictions at Anthropic, a U.S. company rather than a Chinese one.
"You're easing export controls to China, you're leaving loopholes, and the first time you use export controls on artificial intelligence is to hit an American company?" McGuire said.
