What happened
The dispute between the Pentagon and Anthropic over permitted military uses of the Claude AI system has, according to the sources, grown beyond a contract negotiation. Anthropic said on February 27 that Defense Secretary Pete Hegseth had directed the department to designate the company a supply-chain risk after it refused to remove restrictions on mass domestic surveillance of Americans and fully autonomous weapons. The important distinction: what is documented is a direction to pursue a designation, not a published final designation or order. Anthropic said it would challenge any such action in court.
The deadline and background
ABC News reported that Pentagon officials had given Anthropic a Friday deadline to accept broader military-use terms, or risk losing Pentagon business and facing further measures. According to ABC, Anthropic was participating in a Pentagon pilot program valued at $200 million. CBS News reported that officials sent what was described as a "best-and-final offer" with a Friday deadline. CBS also reported that Anthropic received a $200 million Pentagon contract in July. These figures describe a meaningful but limited direct commercial relationship; the larger question is whether restrictions could constrain other defense-sector deployments of the model.
According to Anthropic, it does not dispute that Claude should be available to national-security agencies. The company said it supports lawful national-security uses apart from the two stated exceptions. It said that, to the best of its knowledge, its policies had not interfered with any government mission. Pentagon officials described the practical risk differently: the department wants lawful use, the end user is responsible for legality, and it does not intend to use the tools for domestic surveillance or autonomous weapons.
Point of contention: who enforces the limits?
According to the sources, the core of the dispute is less about a disclosed operational plan than about the contractual question of who enforces use limits: the AI supplier or the government end user. Anthropic argues the latest proposed language made virtually no progress on its two restrictions and could allow its safeguards to be disregarded. Relying on an end user's legal obligations is insufficient, it argues, when the system could be directed into sensitive intelligence or weapons workflows. The Pentagon places the operative compliance burden on the deploying agency or contractor. Anthropic wants the use boundaries embedded in the terms under which the technology is made available.
Unclear: scope of a designation
Anthropic said Hegseth had suggested that companies doing business with the military could be restricted from doing business with Anthropic. The company contends that a designation under 10 U.S.C. 3252 could apply only to Claude use in Department of War contract work, not to commercial customers or to contractors' work outside department contracts. This is Anthropic's legal interpretation, not a settled description of a final action. Geoffrey Gertz of the Center for a New American Security told NPR that such tools have traditionally been aimed at technology tied to foreign adversaries.
Defense Production Act: option, not confirmed
Reporting also raised the possibility of using the Defense Production Act to compel access. ABC News described this as a possibility, and CBS reported that officials were considering it. Neither account says the law has actually been invoked.
Assessment
For German companies that handle US defense contracts or deploy frontier models in regulated settings, the case illustrates how much contractually fixed use limits matter. If a designation with broad reach were actually issued, compliance and legal reviews could tighten for suppliers and subcontractors before a court has decided on the authority. Whether that happens remains open. The next consequential step is likely to be a formal department document specifying the technology, entities, contracts and uses the action would cover.
Sources
Editorially owned by Ideal Syka. Sources and method: Newsroom & method. Tips and corrections: ai@i6eal.de.
