在 2026-03-20（20:03）提交的法庭文件中，Anthropic 公部门主管 Thiyagu Ramasamy 主张，一旦美军在其系统中运行生成式模型 Claude，Anthropic 「从未」也「没有」能力使其停止运作、改变功能、关闭存取或以其他方式影响前线行动；他并明言不存在远端「后门」或「kill switch」，公司人员也无法登入国防部系统于作战中修改或停用模型。

五角大厦与该 AI 实验室已争论数月，国防部长 Pete Hegseth 本月将 Anthropic 定为供应链风险，导致国防部在未来数月不得（含承包商）使用其软体，其他联邦机构亦在放弃 Claude。Anthropic 已提出 2 起诉讼，主张禁令违宪并申请紧急救济，但客户已开始取消交易；其中一案订于 2026-03-24 在旧金山联邦地方法院开庭，法官可能随后不久决定是否暂时撤销。

政府律师称国防部「无须容忍」关键军事系统在国防与现役作战的关键时刻受损之风险；其论点是，若 Anthropic 不同意特定用途，可能透过关闭 Claude 存取或推送有害更新来干扰作战。Ramasamy 反驳称，任何更新需经政府与云端供应商（文件暗示为 Amazon Web Services）同意，且 Anthropic 无法存取军方输入的 prompts 或资料；政策主管 Sarah Heck 亦在 2026-03-20 的文件中称，公司于 2026-03-04 的合约草案愿保证不控制或否决国防部（文件亦称 Department of War）之合法作战决策，但谈判最终破局；同时国防部表示正与第三方云端业者合作，采取额外措施以确保 Anthropic 领导层无法单方面变更现行 Claude 系统。

In a court filing dated 2026-03-20 (20:03), Anthropic public-sector head Thiyagu Ramasamy said that once the US military runs the generative model Claude on its own systems, Anthropic has never had—and does not have—the ability to make it stop working, alter functionality, shut off access, or otherwise affect operations; he added there is no remote back door or “kill switch,” and company personnel cannot log into Defense Department systems to modify or disable models during a mission.

After months of disputes over national-security use, defense secretary Pete Hegseth labeled Anthropic a supply-chain risk this month, a designation that will bar the Department of Defense from using its software (including via contractors) over the coming months, while other federal agencies are also abandoning Claude. Anthropic has filed 2 lawsuits challenging the ban’s constitutionality and seeking emergency relief, but customers have already begun canceling deals; a hearing in one case is set for 2026-03-24 in federal district court in San Francisco, and the judge could decide on a temporary reversal soon after.

Government attorneys argued the Pentagon “is not required to tolerate” the risk of critical military systems being jeopardized at pivotal moments for national defense and active operations, asserting Anthropic could disrupt wartime use by cutting off Claude access or pushing harmful updates if it disapproved of certain applications. Ramasamy countered that updates would require approval from the government and the cloud provider (implicitly Amazon Web Services) and that Anthropic cannot access military users’ prompts or data; policy head Sarah Heck said a 2026-03-04 contract proposal would have guaranteed Anthropic no right to control or veto lawful operational decision-making (the filings also used “Department of War”), but talks ultimately collapsed; meanwhile, the Defense Department says it is taking additional mitigation measures with third-party cloud providers to ensure Anthropic leadership cannot make unilateral changes to deployed Claude systems.