美国国防部与人工智能公司Anthropic之间的争议突显了AI在军事系统中的快速扩展。冲突焦点在于政府要求对Claude模型的无条件访问，而Anthropic拒绝将其用于大规模国内监控或完全自主武器系统。随后五角大楼将其产品标记为“供应链风险”，促使该公司在2026年3月提起两项诉讼。与此同时，Palantir在2024年11月宣布将Claude整合到其面向情报和国防机构的软件中，引发外界对AI在军事行动中角色的关注。据报道，该技术已在海外行动中使用，并在2026年1月协助一次行动，导致委内瑞拉总统尼古拉斯·马杜罗被捕。

Palantir在美国军事AI生态中占据关键地位。自2017年以来，该公司一直是国防部“Project Maven”（算法战争跨职能团队）的主要承包商。Maven由国家地理空间情报局管理，可被陆军、空军、太空军、海军、海军陆战队和中央司令部等多个机构访问，并正在“整个部门”部署。系统利用计算机视觉算法分析卫星等“太空资产”的图像，可自动识别潜在敌方系统、区分人员与车辆，并可可视化目标。工具还可“提名”打击目标，并通过“AI Asset Tasking Recommender”为目标分配轰炸机和弹药，同时在军方之间传递目标情报和战场态势报告。

Palantir的Artificial Intelligence Platform（AIP）通过聊天机器人接口将大型语言模型嵌入现有系统（如Foundry或Gotham）。用户可选择来自Anthropic、Google或Meta的模型，并指定训练数据来源。演示显示，一名军事分析员可通过聊天界面请求情报解释并生成三种作战方案，例如空中打击、远程炮兵或战术小队。系统还能在数秒内生成部队行进路线并分配电子干扰设备。另一个案例显示模型生成五种战略选项，并可分析卫星图像。企业版本Claude还可在约200词报告中总结军事行动影响，将原本可能需要约5小时研究的情报分析自动化。

A dispute between the US Department of Defense and AI company Anthropic highlights the rapid expansion of artificial intelligence in military systems. The conflict centers on government demands for unconditional access to Claude models, which Anthropic refused due to concerns about mass domestic surveillance and fully autonomous weapons. The Pentagon subsequently labeled the company’s products a “supply-chain risk,” prompting two lawsuits filed in March 2026. Meanwhile, Palantir announced in November 2024 that it would integrate Claude into software used by intelligence and defense agencies, increasing scrutiny of AI’s operational role. The technology has reportedly been used in overseas operations and in January 2026 helped support an operation leading to the capture of Venezuelan president Nicolás Maduro.

Palantir occupies a central position in the US military AI ecosystem. Since 2017 it has served as the primary contractor for the Defense Department’s Project Maven, the Algorithmic Warfare Cross-Functional Team. Managed by the National Geospatial-Intelligence Agency, Maven can be accessed by the Army, Air Force, Space Force, Navy, Marine Corps, and US Central Command and is reportedly being deployed “across the entire department.” The system uses computer vision algorithms to analyze satellite and other “space-based asset” imagery, automatically detecting likely enemy systems and distinguishing people from vehicles. It can visualize targets, nominate them for strikes, and use an “AI Asset Tasking Recommender” to assign bombers and munitions while transmitting target intelligence and battlefield situation reports.

Palantir’s Artificial Intelligence Platform (AIP) embeds large language models through a chatbot interface within existing systems such as Foundry or Gotham. Users can select models from Anthropic, Google, or Meta and specify which datasets the system draws upon. Demonstrations show a military analyst querying the chatbot to interpret intelligence and generate three operational options, including air strikes, long-range artillery, or tactical ground teams. The system can also generate troop routes and assign electronic jamming assets within seconds. Another example shows the model producing five strategic options and interpreting satellite imagery. Enterprise Claude can also generate intelligence assessments, including a 200-word summary of military operations, automating analysis that previously might require roughly five hours of research.