UDPATE 2/25, 11:30 A.M.: Anthropic has since scaled back its commitment to its safety guardrails, citing the need for competitiveness in the AI sector.
The Pentagon is demanding Anthropic drop its AI guardrails regarding its use in autonomous weapons systems and surveillance by Friday, or lose a $200 million contract, in what is now a full force battle over the future of AI’s use in the U.S. military.
Anthropic says that its AI technologies should not be used for completely autonomous military targeting or to “spy on Americans en masse.” But the Pentagon says it must be able to use Anthropic’s model for “all lawful use cases" — without limit.
Defense Secretary Pete Hegseth warned that if Anthropic does not back down, the Pentagon will render Anthropic’s AI model, Claude, a “supply chain risk” giving it a black mark for other contracts.
In an apparent public jab at Anthropic back in January, Hegseth said, “We will not employ AI models that won’t allow you to fight wars.”
DoD officials have admitted, however, that gutting and replacing Claude within its current operations would be difficult.
Today’s ultimatum follows a meeting between DoD Secretary Pete Hegseth and Anthropic CEO Dario Amodei, where Amodei held to stated guardrails of how the company wants its tech to be used.
As part of a collaboration Anthropic has with Palantir, the U.S. military used Anthropic’s Claude model to prepare for an operation that removed Venezuelan leader Nicolas Maduro from power last month. Subsequent disagreement with Palantir over Claude’s use in the Maduro raid has led to a greater rift with the Pentagon.
Anthropic’s competitors have not all held to its guardrails. Yesterday, xAI agreed that its tech could now be used for classified purposes, under any lawful use by the DoD.
OpenAI, Anthropic, Google, and xAI all received $200 million DoD contracts in July 2025 to “develop prototype frontier AI capabilities to address critical national security challenges in both warfighting and enterprise domains.” Of those four companies, Anthropic was the first to be approved and deployed for use in classified settings, where it can work with other partners like Palantir as part of that approval.
