Anthropic is holding its line in a high-stakes dispute with the Pentagon over how its AI systems can be used by the military. CEO Dario Amodei said Thursday that he “cannot in good conscience accede to [the Pentagon’s] request” for unrestricted access.
The disagreement is not about whether Anthropic can work with the military at all. According to the source article, the company’s position is narrower: it wants two specific limits to remain in place before its technology is used in the most sensitive defense settings.
The two safeguards at the center of the fight
Amodei framed the issue as a boundary around uses that Anthropic believes could create risks to democratic values or exceed what current AI can safely do. In his statement, he wrote: “Anthropic understands that the Department of War, not private companies, makes military decisions.”
But he also argued that there are cases where the company should not simply remove all restrictions. The two cases named in the article are:
- Mass surveillance of Americans.
- Fully autonomous weapons with no human in the loop.
That makes the conflict unusually specific. Anthropic is not described as trying to control every military use of Claude. The company is objecting to contract language that, in its view, would fail to preserve those two safeguards.
An Anthropic spokesperson told TechCrunch that the contract language received overnight from the Department of War “made virtually no progress” on preventing those uses. The spokesperson also said language presented as compromise was tied to legal wording that would allow the safeguards to be disregarded.
The Pentagon’s position is broader control
The Pentagon’s position, as described in the source, is that it should be able to use Anthropic’s model for all lawful purposes. It also argues that a private company should not dictate how the military uses technology once that technology is embedded in defense work.
Senior Official Jeremy Lewin expressed that broader view in a public post. He wrote that the issue was not Anthropic or the specific disputed terms, but the principle that technology deeply embedded in the military must remain under the exclusive control of duly elected or appointed leaders.
That is the core institutional tension. Anthropic is saying there are narrow AI uses it will not enable. The Pentagon is saying lawful military use should not be governed by a private company’s terms.
Both positions create practical consequences. If Anthropic refuses unrestricted access, it may lose the Department as a customer. If the Pentagon accepts the safeguards, it would be allowing a contractor to keep explicit limits around some uses of a frontier AI system.
A deadline, negotiations, and pressure tactics
Amodei’s statement arrived less than 24 hours before a Friday 5:01 p.m. deadline set by Defense Secretary Pete Hegseth. The Department had given Anthropic a choice: agree to the demands or face consequences.
Even so, Anthropic has not said it is walking away. A spokesperson told TechCrunch that Amodei’s statement does not mean the company is ending negotiations, and that it is continuing to engage in good faith with the Department going forward.
The pressure on Anthropic has included two possible government moves. The Department of Defense has attempted to force Amodei’s hand by either labeling Anthropic a supply chain risk or invoking the Defense Production Act. The source describes the supply chain risk label as a designation reserved for foreign adversaries, and notes that the Defense Production Act gives the president authority to force companies to prioritize or expand production for national defense.
Amodei pointed to what he saw as a contradiction between those threats. As he put it: “One labels us a security risk; the other labels Claude as essential to national security.”
That line captures the unusual shape of the dispute. The same technology is being treated, in the reported government pressure campaign, as both a potential risk and something valuable enough to compel.
What happens if Anthropic is offboarded
Anthropic is currently described in the source as the only frontier AI lab with classified-ready systems for the military. The article also says the DOD is reportedly getting xAI ready for the job.
Amodei acknowledged that the Department has the right to choose contractors aligned with its vision. But he also said Anthropic hopes the Department reconsiders, given the value the company believes its technology provides to the armed forces.
The company’s preferred outcome is to keep working with the Department while preserving the two requested safeguards. Amodei said: “Our strong preference is to continue to serve the Department and our warfighters—with our two requested safeguards in place.”
He also described a fallback path if the Department decides to remove Anthropic from the work. In that case, he said the company would help enable a smooth transition to another provider, with the aim of avoiding disruption to ongoing military planning, operations, or other critical missions.
That matters because the dispute is no longer just a policy debate. It is also an operational question about continuity, classified-ready AI systems, and how far a private AI lab can go in setting boundaries when its models are used by the military.
The larger AI governance question
The source article makes clear that the argument is bigger than one contract clause. It raises a central question for frontier AI: when a system becomes useful enough for national defense, who gets to define the limits?
Anthropic’s answer is that some uses remain outside its acceptable boundaries, even if the government sees them as lawful. The Pentagon’s answer is that military decisions belong to public officials, not private companies.
Neither side is described as rejecting military AI outright. The dispute is about where the final authority sits when AI capabilities intersect with surveillance, weapons, and national security operations.
For now, Anthropic is still at the table. But Amodei’s statement makes clear that the company is prepared to part ways rather than provide unrestricted access without the safeguards it has named.