ANALYSIS:
Anthropic co-founder and CEO Dario Amodei did his duty at President Trump’s White House tech titan summit amid the mogul’s bitter battle with the Pentagon.
Mr. Amodei, who runs Anthropic and its Claude AI model, wanted to decide when in war the military could and could not turn to his brainy model. Defense Secretary Pete Hegseth essentially said commanders, not Mr. Amodei or Claude, run the military.
In Silicon Valley’s AI revolution, Anthropic took a clear lead over other firms represented in the assembled White House cohort, including Elon Musk (xAI), Mark Zuckerberg (Meta) and Jeff Bezos (Amazon founder). Mr. Amodei seeks a $2 trillion valuation when his colossus goes public, press reports say.
People were interested in these podcasts
As Anthropic expands its power and wealth, it is flexing its political muscle. The company and its employees have given $10 million to candidates and political action committees since the company’s 2021 founding, $9 million of that this year alone, according to a review by The Washington Times of Federal Election Commission files. The vast majority of candidate donations go to Democrats.
At the White House on Tuesday, Mr. Amodei reiterated his concerns about rogue models causing harm. One of his researchers told news outlets that he quit for fear that AI could cause human extinction.
Mr. Amodei added his signature to a voluntary joint statement that stresses self-regulation. He also agreed that the U.S. must win the global AI competition, meaning beat China.
Mr. Trump called Mr. Amodei “fantastic” in a White House driveway press conference. “We all need to work together,” Mr. Amodei said.
For months, however, Mr. Amodei has been in a bare-knuckle legal face-off with the administration, which views his team as a possible national security threat.
Mr. Amodei sued the Defense Department and other Cabinet members earlier this year after Mr. Hegseth evicted Claude. Mr. Hegseth called the software a risk to warfighters, not because it might turn villain, but because of people such as Mr. Amodei who invented and operate Claude.
Emil Michael, the Pentagon’s undersecretary of defense for research and engineering, told the story in a declaration filed in the U.S. District Court for the Northern District of California.
Mr. Michael accused Anthropic leaders of “adversarial behavior” and “bad faith,” citing the “behavior of its leaders during talks” to extend their Pentagon contract.
“Anthropic’s demonstrated willingness to interfere with that chain of command is a significant risk,” Mr. Michael stated.
Anthropic does not want its model used to autonomously launch weapons or to surveil Americans. Mr. Hegseth said no weapon is used without an OK from a human operator and that the Pentagon does not spy domestically.
Here are some of Mr. Michael’s blunt criticisms of Anthropic and why it is a risk:
• “The company’s leadership insisted on imposing restrictions on DoD’s lawful military capability development, operations, and intelligence missions, even though it would impair the capabilities of the U.S. military relative to our adversaries. In short, Anthropic made clear that it will not allow the Government to deploy Claude for multiple lawful uses.”
• “It is reasonably likely that Anthropic’s leadership would alter or even shut off DoD’s use of Claude if Anthropic believes that the model may be used for purposes it deems, in its sole discretion, to extend beyond the company’s unilaterally imposed boundaries before or during a military operation, which could endanger the lives of U.S. military personnel and civilians and compromise the United States’ warfighting mission.”
• “The company’s leadership demonstrated bad faith by sharing with the press unclassified but sensitive details of private conversations with DoD leadership in order to exert public pressure on DoD to concede to Anthropic’s demands.”
Mr. Michael also raised concerns about Anthropic’s dealings with China, which he described as a significant risk given the volume of intelligence it collects on America.
Mr. Michael said, “Anthropic employs a large number of foreign nationals to build and support its [large language model] products, including many from the People’s Republic of China (PRC), which increases the degree of adversarial risk should those employees comply with the PRC’s National Intelligence Law. Although other major U.S. AI labs that provide LLM products to DoD may present similar risks, the technical and security assurances of the other labs’ leadership, along with their consistently responsible and trustworthy behavior during their engagement with DoD, mitigate these risks. Anthropic’s case, however, is different.”
The Justice Department’s court filing was just as blunt.
“AI systems are acutely vulnerable to manipulation, and Anthropic could attempt to disable its technology or preemptively alter the behavior of its model either before or during ongoing war fighting operations, if Anthropic — in its discretion — feels that its corporate ’red lines’ are being crossed.”
The filing disclosed that Anthropic has already objected to certain uses of Claude, presumably in the war against Iran.
Such events led the Pentagon to conclude it “cannot trust Anthropic to ensure the integrity of its models.”
The so-called doomers on social media express fears of an out-of-control AI destroying the world. The Trump team fears out-of-control Anthropic humans could cause “disastrous action that would put our warfighters’ lives in danger.”
The Times’s review of FEC records shows that the vast majority of Anthropic employees’ donations go to Democrats and PACs such as ActBlue. A Google AI analysis of FEC data puts the 2026 candidate ratio at 68.8% Democrats and 12.9% Republicans.
Daniela Amodei, Anthropic president, co-founder and sister of Dario, contributed to the Kamala Harris presidential campaign.
In the court battle, U.S. District Judge Rita F. Lin, a Biden appointee, sided with Anthropic in August.
On Sept. 25, however, the U.S. Court of Appeals for the District of Columbia Circuit backed the Pentagon’s right to blacklist Anthropic.
Contact the author
Please read our comment policy before commenting.