A contentious and escalating dispute between the U.S. Department of Defense and artificial intelligence firm Anthropic has cast a spotlight on the intricate and often opaque ways generative AI is being integrated into American military operations. At the heart of the conflict lies Anthropic’s refusal to grant the Pentagon unfettered access to its Claude AI models, citing critical ethical concerns. The AI developer has stipulated that its technology should not be employed for mass surveillance of American citizens or for the development of fully autonomous weapons systems. In response to these conditions, the Pentagon, under the Trump administration, designated Anthropic’s products as a "supply-chain risk." This designation has prompted Anthropic to take legal action, filing two lawsuits this week alleging unlawful retaliation and seeking to overturn the Pentagon’s decision.
The public friction between the defense establishment and a leading AI innovator is occurring against the backdrop of a rapidly intensifying geopolitical climate, particularly the escalating conflict in Iran. This confluence of events has amplified scrutiny of Anthropic’s strategic partnership with Palantir Technologies, a prominent defense contractor. In November 2024, Palantir announced its intention to integrate Claude AI models into the software solutions it provides to U.S. intelligence and defense agencies. Palantir has asserted that this integration aims to empower analysts with enhanced capabilities to uncover "data-driven insights," identify complex patterns, and facilitate "informed decisions in time-sensitive situations."
Despite these pronouncements, concrete details regarding the specific applications of Claude within the military and the particular Pentagon systems that leverage its capabilities remain scarce. This lack of transparency persists even as reports suggest that Claude is actively being utilized in certain U.S. defense operations abroad, including those related to the conflict in Iran. Earlier this year, in January, Claude reportedly played a pivotal role in a U.S. military operation that led to the apprehension of Venezuelan President Nicolás Maduro.
A comprehensive review by WIRED of Palantir software demonstrations, publicly available documentation, and Pentagon records offers the most detailed insight to date into how U.S. military officials might be deploying AI chatbots. This review sheds light on the types of queries being submitted to these systems, the data sources employed for generating responses, and the nature of the recommendations provided to analysts.
When contacted for comment, the Department of Defense did not respond. Representatives for both Palantir and Anthropic declined to comment on the ongoing situation.
Palantir’s Deep-Rooted Pentagon Ties and the Genesis of AI Integration
The U.S. military’s engagement with advanced AI capabilities for intelligence analysis has been a developing narrative for years. Military personnel can reportedly leverage Claude to process and analyze vast quantities of intelligence data. Palantir, a key player in this ecosystem, offers multiple software platforms to the Pentagon where such analytical tasks could be performed. However, the company has not publicly specified which of these platforms incorporate Claude’s functionalities.
A significant milestone in the military’s AI journey was Project Maven, initiated in 2017. Palantir has served as the primary contractor for this initiative, formally known as the Algorithmic Warfare Cross-Functional team, a Defense Department endeavor focused on deploying artificial intelligence in combat scenarios. For Project Maven, Palantir developed the Maven Smart System, often referred to simply as Maven.
The Maven system is managed by the National Geospatial-Intelligence Agency (NGA), the government body responsible for the collection and analysis of satellite imagery. A broad spectrum of military agencies, including the Army, Air Force, Space Force, Navy, Marine Corps, and U.S. Central Command (which oversees operations in Iran), have access to Maven. In a recent Palantir conference, Cameron Stanley, the Pentagon’s Chief Digital and Artificial Intelligence Officer, indicated that Maven is being deployed "across the entire department."
Publicly available assessments of Maven highlight its capacity to apply "computer vision algorithms" to imagery captured by "space-based assets" such as satellites. The system can automatically detect objects that are likely to be "enemy systems." Demonstrations of Maven have showcased its ability to differentiate between individuals and vehicles.
Further functionalities within Maven are designed to visualize "potential targets" and "nominate" them for aerial or ground-based bombardment. A tool named the AI Asset Tasking Recommender, as demonstrated by Stanley, can propose specific bombers and munitions for designated targets. Maven also facilitates the secure transmission of "target intelligence data and enemy situation reports" among military personnel.
While The New York Times and The Washington Post have reported that Maven utilizes Anthropic’s AI technology, WIRED has not independently verified these claims.
The Army Intelligence Data Platform and Palantir’s Broader AI Suite
Beyond Project Maven, Palantir has also supplied another critical intelligence platform to the U.S. Army since 2022: the Army Intelligence Data Platform (AIDP). Palantir has stated that the AIDP "integrates" data from Maven and at least four other government systems. Details about the AIDP remain limited in the public domain, but military assessments have described it as a tool capable of preparing intelligence briefings prior to military operations and graphically depicting the positions of troops and weaponry. It also includes a feature called Dossier, reportedly used for developing an "intelligence running estimate"—a continuously updated compilation of battlefield information that precedes a final intelligence summary. The extent to which Claude is integrated into Palantir’s AIDP remains unclear.
Although Palantir has not disclosed which of its Pentagon systems are equipped with Claude, the company has provided some insights into its potential integration. In its November 2024 press release announcing the partnership with Anthropic, Palantir noted that Claude had become accessible within the Artificial Intelligence Platform (AIP), one of Palantir’s newer commercial offerings, earlier that month.
Unpacking Palantir’s Artificial Intelligence Platform (AIP)
Palantir’s Artificial Intelligence Platform (AIP) is not a standalone product but rather an application designed to function within existing Palantir systems like Foundry or Gotham. AIP offers users a chatbot interface, referred to by the company as an AIP Assistant or AIP Agent. This assistant can respond to queries and execute tasks within the broader system.
The AIP Assistants are powered by third-party large language models (LLMs) from various providers, including Anthropic, Google, and Meta. Customers have the flexibility to select their preferred LLM and specify the training data used for generating responses. This feature is particularly significant for intelligence and national security applications, where data classification is a critical consideration.
A Palantir demonstration released in 2023 illustrates how an AIP Assistant could aid a "military operator responsible for monitoring activity within Eastern Europe" in planning and authorizing a ground attack on tanks through a conversational interface with the chatbot. The process commences with an automated alert from the AIP Assistant regarding "potential unusual enemy activity," detected through AI processing of radar imagery. In this instance, a computer vision algorithm, rather than an LLM, would likely identify the anomalous activity. The AIP Assistant then assists the analyst in interpreting the findings and determining the appropriate course of action. While the chatbot may not directly identify targets, its role in guiding the analyst’s decision-making process could be instrumental in transforming a nascent observation into a concrete action.
When prompted with the question, "What enemy military unit is in the region?", the AIP Assistant might infer, "likely an armor attack battalion based on the pattern of the equipment." This suggestion could prompt the analyst to request surveillance via an MQ-9 Reaper drone. Subsequently, the analyst could ask the AIP Assistant to "generate three courses of action to target this enemy equipment." The assistant would then propose options such as engaging the unit with an "air asset," "long-range artillery," or a "tactical team." In the demonstrated scenario, the user directs the assistant to forward these options to a hypothetical commander, who ultimately selects the tactical team.
The subsequent stages unfold rapidly. The analyst instructs the AIP Assistant to "analyze the battlefield," "generate a route" for troop deployment to intercept the enemy, and finally, "assign jammers" to disrupt their communications. Within seconds, the analyst reviews the battle plan and authorizes troop mobilization.
In this specific scenario, Claude would function as the "voice" of the AIP Assistant, providing the reasoning behind its generated responses. Other AIP demonstrations showcase users engaging with LLMs in similar conversational formats. A recent Palantir blog post detailed how NATO, a customer of Maven Smart Systems, could utilize an AIP Agent within the tool. The graphic depicts a third-party defense contractor selecting from various AI models, including different versions of OpenAI’s ChatGPT and Meta’s Llama. While the example shows the selection of GPT 4.1, it implies that a soldier could similarly opt for Claude.
An analyst then views a digital map displaying troop and weapon locations. Within a "COA" (courses of action) panel, a button click initiates a tool powered by GPT-4.1 to generate five potential military strategies, including one titled "Support-by-Fire-Then-Penetration-Shock-and-Destruction." Another example illustrates the system’s capability to interpret satellite imagery. The analyst selects three tanker truck detections on a map, inputs them into the AIP Agent’s chat interface, and requests it to "interpret" the imagery and suggest subsequent actions.
Claude may also be instrumental in the creation of intelligence assessments that inform future strike planning. In June 2025, a demonstration by Kunaal Sharma, Anthropic’s public sector lead, showcased how the enterprise version of Claude could generate detailed reports on a Ukrainian drone strike, dubbed "Operation Spider’s Web." Sharma highlighted that in this demonstration, Claude relied solely on publicly available information. However, through the partnership with Palantir, he explained, federal agencies can integrate internal datasets. Sharma elaborated that generating such comprehensive reports, which previously required extensive manual research and analysis, could be significantly expedited by AI.
In the demo, Sharma tasked Claude with creating an "interactive dashboard" of information pertaining to Operation Spider’s Web, and subsequently translating it into "object types" amenable to analysis within Palantir’s Foundry software. He also requested Claude to produce a detailed analysis of recent developments in Russia’s border provinces, alongside a 200-word synopsis of the operation’s "military and political effects." Sharma remarked on the quality of the AI-generated analysis, stating it was comparable to human-authored reports he had reviewed over two decades in academia and intelligence.
Broader Implications and Unanswered Questions
The dispute between the Pentagon and Anthropic underscores a critical tension in the rapid adoption of advanced AI by military organizations: the balance between operational necessity and ethical safeguards. Anthropic’s stance reflects a growing concern within the AI development community about the potential misuse of powerful AI technologies, particularly in contexts with profound implications for human rights and international stability. The company’s insistence on restrictions against mass surveillance and fully autonomous weapons suggests a commitment to responsible AI deployment, even at the risk of alienating a significant governmental client.
The Pentagon’s "supply-chain risk" designation, while potentially a bureaucratic maneuver to exert control, also highlights the vulnerabilities inherent in relying on third-party AI technologies. The military’s dependence on external developers for cutting-edge AI capabilities raises questions about long-term control, data security, and the ability to independently audit and verify the ethical compliance of these systems.
The legal challenges initiated by Anthropic are likely to have far-reaching consequences, potentially setting precedents for how AI developers can challenge governmental actions and how the military can procure and deploy AI technologies. The outcomes of these lawsuits could influence future procurement policies, contractual agreements, and the ethical frameworks governing the use of AI in defense.
Furthermore, the lack of transparency surrounding the specific applications of Claude within military operations, as detailed by WIRED’s investigation, fuels public concern. In an era where AI is increasingly shaping geopolitical dynamics, understanding the scope and nature of its deployment in warfare is paramount for democratic oversight and accountability. The current opacity surrounding these advanced AI integrations leaves critical questions unanswered about the potential for unintended consequences, algorithmic bias, and the erosion of human control in critical decision-making processes. The ongoing legal battle and the geopolitical tensions serve as a stark reminder of the complex ethical and operational challenges that lie ahead as the world navigates the integration of artificial intelligence into the fabric of national security.