The Pentagon and Anthropic, a prominent AI developer, are locked in a dispute after Anthropic’s chief executive rejected a Defense Department demand to relax safety restrictions on the company’s model or face exclusion from military work. The disagreement could affect existing contracts worth up to about $200 million and access to advanced AI tools for defense applications.
Why they disagree
Anthropic CEO Dario Amodei has repeatedly said his company will not allow its model, Claude, to be used for domestic mass surveillance or to power fully autonomous weapons that can kill without human authorization. Amodei has called those uses “bright red lines” and said they are illegitimate.
Pentagon officials counter that they do not plan to use Anthropic’s systems for mass surveillance or to deploy autonomous weapons without human oversight. However, they argue companies should not unilaterally decide which lawful applications the government may pursue. A senior Defense Department official told reporters that determining whether a use is legal is the Pentagon’s responsibility as the end user, and contractors should permit government use “for all lawful purposes.”
Amodei’s response and recent exchanges
Anthropic said it could not accept the department’s revised contract language, which the company believes fails to prevent Claude from being used in domestic surveillance or in fully autonomous weapons. Anthropic added that some of the proposed concessions were paired with legal wording that would allow those safeguards to be overridden. The company said it remains open to talks and prioritized uninterrupted support for the Department and U.S. service members, but maintained the narrow prohibitions are necessary because current technology cannot safely handle those uses.
Relations have become strained. Reportedly, at a meeting between Defense Secretary Pete Hegseth and Amodei, Hegseth warned of consequences if Anthropic did not comply. One person familiar with the discussions said Hegseth suggested canceling Anthropic’s roughly $200 million Pentagon contract. A Pentagon official also said the department could seek remedies including forcing Anthropic to permit government use contrary to the company’s preferences or effectively blacklisting the company from future defense contracts.
A hard deadline
Pentagon spokesman Sean Parnell posted that Anthropic had until 5:01 p.m. ET on Friday to accept the department’s terms before the Pentagon would move forward. Parnell framed the demand as intended to avoid any impression that the department is seeking to use AI for surveillance or fully autonomous weapons, reiterating that the department does not support those applications.
Possible Pentagon actions and legal questions
The department has warned it could label Anthropic a “supply chain risk,” a designation more commonly applied to technology from foreign adversaries. Experts say the practical effects of such a label are unclear: it might bar other defense contractors from using Anthropic’s tools or it could lead to broader prohibitions, which would be especially consequential.
Officials have also suggested they could use the Defense Production Act (DPA) to compel Anthropic to remove guardrails. Legal and policy scholars call that a rare and unconventional use of the DPA, which is usually reserved for urgent national emergencies to direct industrial production or allocation of resources. Forcing a private company to alter product safeguards under the DPA would be an unusual application and would likely trigger legal challenges.
Geoffrey Gertz, a senior fellow at the Center for a New American Security, noted a tension in the department’s posture: officials are both saying Anthropic is too risky to remain in the supply chain while also implying the company may be so indispensable they might need to compel it to comply. If the department simply cancels the contract, the dispute could end there. But attempts to force guardrail removal or impose a sweeping supply-chain designation would probably provoke litigation.
Wider implications
The contested contract represents a relatively small portion of Anthropic’s reported revenue, but the outcome could have outsized importance. Anthropic was the first company cleared for classified use after its model was judged secure and capable for sensitive applications. How this standoff resolves could influence whether and how private AI companies can set limits on military uses of their products, how much authority the U.S. government has to override company-imposed safety measures, and whether legal and policy battles will follow.
Observers expect that if the department escalates its approach, the fight will move into court and fuel broader debates over the balance between national security needs and safeguards intended to prevent dangerous uses of AI.