Pentagon Gives A.I. Company an Ultimatum

 

Pentagon Gives A.I. Company an Ultimatum

 

Anthropic insists on limits on how its technology is used and could be labeled a supply chain risk if it fails to accept the military’s demands.

 

Julian E. Barnes Sheera Frenkel

By Julian E. Barnes and Sheera Frenkel

Julian E. Barnes reported from Washington, and Sheera Frenkel from San Francisco.

https://www.nytimes.com/2026/02/24/us/politics/pentagon-anthropic.html

Feb. 24, 2026

 

The Pentagon delivered an ultimatum to Anthropic, the only artificial intelligence company currently operating on classified military systems, ordering the firm to bend to its demands by Friday.

 

If the firm fails to agree by 5:01 p.m. on Friday, Defense Secretary Pete Hegseth said the Trump administration would invoke the Defense Production Act, compelling the use of its model by the military and labeling the company a supply chain risk, according to a senior Pentagon official. That step would put Anthropic’s government contracts at risk.

 

The two threats are fundamentally at odds: One would prevent the government from using the company’s products, while the other would force the company to let the government use the products.

 

Despite the contradiction, the threats reflect the level of anger in the top ranks of the Pentagon toward Anthropic for resisting its demands and how important the company’s model has become to the military.

 

“The Pentagon knows they are issuing an extreme threat. They are using every button or lever they have,” said Jessica Tillipman, an associate dean at the George Washington University Law School. “The bigger issue here is that it waters down these designations. They are transforming what is designed to be national security tools into a point of leverage for business.”

 

Mr. Hegseth summoned Dario Amodei, the Anthropic chief executive, to the Pentagon on Tuesday for a morning meeting. The tone of the discussion was civil, but when Anthropic did not agree to Mr. Hegseth’s demands, he leveled the threats against it, according to people briefed on the meeting.

 

The New York Times spoke to people on both sides of the debate over Anthropic’s work with the military, but they spoke on the condition that their names not be used to discuss the sensitive negotiations.

 

Anthropic has argued that it was asking for reasonable assurances that its model would not be used for surveillance of Americans or in autonomous weapons, such as drone operations, that did not involve human oversight.

 

Anthropic’s supporters have contended that the company is being punished for being first on the classified system and creating a special model, Claude Gov, that does not have the same guardrails and restrictions that their models available to the public have.

 

Pentagon officials have said that using software and weapons lawfully is their responsibility, one they take seriously. But the officials say they cannot effectively allow all their contractors to specify how the equipment they sell to the Pentagon will be used, and that lawful use must be the only constraint.

 

While the Defense Production Act gives the Pentagon wide-ranging powers, it is usually invoked in manufacturing contexts. It would be unusual for the act to be used on a software company, forcing Anthropic to make its product available for free.

 

An Anthropic spokesman said that the company had continued good-faith conversations in the meeting at the Pentagon. The spokesman said the company wanted to support the government but needed to ensure that its models were used in line with what they could “reliably and responsibly do.”

 

But the senior Pentagon official rejected those demands and said the debate had nothing to do with those issues. The Pentagon wants all artificial intelligence contracts to stipulate that the military can use the models for any lawful purpose.

 

The official confirmed that the Pentagon has an agreement with Elon Musk’s company xAI to use its artificial intelligence model, Grok, on the classified system. But it will take time to integrate Grok onto classified cloud servers and into software from Palantir, a data analytics company that the military uses. More important, Anthropic’s Claude is considered a superior product to Grok, regularly yielding more accurate information.

 

The Pentagon also is close to an agreement with Google to bring its Gemini model onto the classified system, but the senior official said the deal was not complete.

 

A person briefed on the meeting said Anthropic would continue to demand assurances that its models are not used for autonomous weapons programs or mass surveillance.

 

Pentagon officials took issue with Anthropic after Palantir reported a conversation that one of its employees had had with a counterpart at the artificial intelligence company regarding the U.S. military operation last month to capture President Nicolás Maduro of Venezuela.

 

In the meeting on Tuesday, Mr. Amodei said there had been a misunderstanding and that his company had not reached out to Palantir or the Pentagon about the Maduro operation, according to a person briefed on the meeting.

 

Mr. Amodei insisted his company had never objected to or interfered with legitimate military operations.

 

A correction was made on Feb. 24, 2026: An earlier version of this article incorrectly described Anthropic’s artificial intelligence model Claude Gov. Its guardrails and restrictions are different from those of the company’s models available to the public, not from classified models.

 

Julian E. Barnes covers the U.S. intelligence agencies and international security matters for The Times. He has written about security issues for more than two decades.

 

Sheera Frenkel is a reporter based in the San Francisco Bay Area, covering the ways technology affects everyday lives with a focus on social media companies, including Facebook, Instagram, Twitter, TikTok, YouTube, Telegram and WhatsApp.