Anthropic on Thursday said it “could not in good conscience” comply with the Pentagon’s request to remove security measures from its artificial intelligence models and give the U.S. military unfettered access to its AI capabilities.
The Pentagon had threatened to terminate the $200 million contract and label Anthropic a “supply chain risk,” a designation with significant financial implications, if Anthropic did not meet its demands by Friday.
CEO Dario Amodei said in a statement that Defense Secretary Pete Hegseth’s threat would not change the company’s position and hoped that Hegseth would “reconsider.”
“Our strong desire is to continue to serve the Department of Defense and the warfighter by taking the two safety measures that have been requested,” he said. “We stand ready to continue our efforts to support the national security of the United States.”
At the heart of the conflict between the Pentagon and Anthropic is a disagreement over how the AI company should be allowed to use its product, Claude. The Pentagon is demanding that Anthropic remove the safety guardrails and allow Claude to be used legally, but Anthropic has balked at allowing Claude to be used in autonomous weapons systems that can kill people without mass domestic surveillance or human intervention.
After months of wrangling and government pressure, Hegseth reportedly gave Amodei until Friday night to agree to the Pentagon’s demands or face punitive action.
Whether Anthropic would concede was seen as a high-profile test of its claims to be the most safety-focused of the big AI companies, as well as whether some in the AI industry would push back against the government’s desire to use the technology for controversial and potentially deadly purposes.
Amodei said in a statement that using AI for autonomous weapons and domestic mass surveillance is “far beyond what can be done safely and reliably with today’s technology.”
The Pentagon has awarded a number of lucrative contracts to tech companies in recent years to build and integrate AI technology into U.S. military systems. Last July, Anthropic was one of several large technology companies, including Google and OpenAI, to sign contracts worth up to $200 million with the Department of Defense. What sets Anthropic apart, and puts it at odds with the Pentagon, is that until this week it was the only AI model approved for use on classified military systems. (Elon Musk’s xAI reached an agreement earlier this week to be used in classified systems as well.)
Anthropic’s technology is reportedly already being used in military applications, including when the US captured Venezuelan leader Nicolas Maduro last month, highlighting the growing use of AI in conflicts. Advances in autonomous weapons technology, such as drones that can carry out operations even after disconnecting from a human operator, are also reinforcing long-standing concerns about how AI might be used in life-or-death situations.
Anthropic and Amodei, long the industry’s most prominent advocates for regulation and safety precautions in AI development, this week watered down their core policy of not releasing new AI models before guaranteeing their safety by striking a deal with the military. Mr. Amodei’s calls for regulation and his history of political opposition to Donald Trump run counter to Mr. Hegseth’s vows to rid the military of “woke men” and pursue an aggressive military policy.
If Mr. Hegseth follows through on his threat to classify Anthropic as a supply chain risk, it would be a major blow to AI companies. The designation is intended more generally for use against foreign adversaries and would prohibit other vendors doing business with the U.S. military from using Anthropic’s products.
