More than 400 pages of government contract documents, obtained through a Freedom of Information Act lawsuit filed with the help of Legal Advocates for Safe Science and Technology, have pulled back the curtain on the relationship between America’s leading artificial intelligence companies and the U.S. Department of Defense. The documents, published by The Intercept, reveal deals signed in July 2025 between the Pentagon and Anthropic, Google, OpenAI, and xAI — each worth up to $200 million — and detail how these companies agreed to build AI prototypes specifically designed to “improve military advantage, military utility, or enhance military decision making” across the entire armed forces.

What the documents make clear is that these arrangements go far beyond routine software procurement. Engineers from frontier AI labs are working directly alongside defense officials to develop systems for surveillance, targeting, logistics, and automated military decision-making. The contracts form the foundation of AI tools now being deployed across the Department of Defense’s classified networks.

The Contracts: What Was Actually Agreed To

In July 2025, the DoD’s Chief Digital and Artificial Intelligence Office (CDAO) announced prototype contract awards of up to $200 million each for Anthropic, Google, OpenAI, and Elon Musk’s xAI. The official language at the time was carefully measured. Pentagon spokesperson Sean Parnell stated that the department must “equip our warfighters with 21st-century technology in order to defeat 21st-century threats.” Google described its role in similarly vague terms, referencing “agentic AI” and “enterprise systems.”

The contract documents obtained by The Intercept tell a more specific story. The AI prototypes were aimed at applications spanning “warfighting” and automated decision-making to military logistics and intelligence operations. An “information sharing” provision embedded in all four agreements created a two-way pipeline between the companies and the Pentagon, establishing an unusually intimate collaboration between private AI developers and the U.S. military apparatus.

Sophia Goodfriend, a University of Cambridge research fellow and non-resident fellow at the Harvard Kennedy School’s Middle East Initiative who studies the impact of big data and machine learning on military conflict, told The Intercept: “This appears to be the first time the extent of collaboration between frontier AI labs and the Pentagon is spelled out in explicit terms. What’s particularly notable is the terms by which engineers are working in lockstep with the department of war to engineer AI systems for surveillance, targeting, and killing.”

The Anthropic Fallout: A Dispute Over Guardrails

Among the four companies involved, Anthropic’s relationship with the Pentagon became the most publicly contentious. The company had been the first major AI firm to have its services cleared for use on classified networks, operating through a partnership with Palantir on Amazon Web Services infrastructure using its Claude Gov models — custom-built for national security customers.

In its July 2025 announcement, Anthropic described its commitment to “responsible AI deployment,” citing “rigorous safety testing, collaborative governance development, and strict usage policies.” The company framed its work with the DoD around developing prototypes fine-tuned on DoD data, collaborating with defense experts on adversarial risk forecasting, and exchanging technical and operational feedback.

Yet behind the scenes, a dispute was forming. According to reporting by The Intercept, the actual falling out between Anthropic and the Pentagon hinged on a specific government contract that had not been made public. The core tension, which eventually played out in public, centered on the Pentagon’s desire for Anthropic to agree that its systems could be used for “any lawful purpose” — language Anthropic reportedly resisted, particularly over concerns related to domestic surveillance and autonomous weapons control.

OpenAI Renegotiates: Surveillance Language Added After Backlash

OpenAI’s path through these negotiations became its own flashpoint. In late February 2026, the company reached a deployment agreement with the Pentagon for use of its systems on classified networks. The initial contract language faced immediate criticism from observers who argued it left open significant loopholes for domestic surveillance. OpenAI CEO Sam Altman subsequently announced a reworked agreement in early March 2026.

The updated contract added explicit language stating: “Consistent with applicable laws, including the Fourth Amendment to the United States Constitution, National Security Act of 1947, FISA Act of 1978, the AI system shall not be intentionally used for domestic surveillance of U.S. persons and nationals.” A second clause clarified that deliberate tracking, surveillance, or monitoring of U.S. persons — including through commercially acquired personal information — was prohibited.

OpenAI also stated publicly that its agreement excluded use by Pentagon intelligence agencies such as the NSA, with Katrina Mulligan, OpenAI’s head of national security partnerships, noting that defense intelligence components were excluded from the contract and that any future NSA work would require a separate agreement with appropriate safeguards.

OpenAI laid out three stated “red lines” governing its DoD work: no use of its technology for mass domestic surveillance, no use to direct autonomous weapons systems, and no use for high-stakes automated decisions without human oversight. The company emphasized a cloud-only deployment architecture in which it retains control of its safety stack and has cleared personnel involved in the process.

The contract language OpenAI shared stated: “The AI System will not be used to independently direct autonomous weapons in any case where law, regulation, or Department policy requires human control, nor will it be used to assume other high-stakes decisions that require approval by a human decisionmaker.”

Critics: Selective Disclosure and Unanswered Questions

Despite the amended language, skepticism persisted among policy watchers and civil liberties advocates. Brad Carson, a former congressman and general counsel of the Army who now leads the Washington, D.C., policy group Americans for Responsible Innovation, pointed to a fundamental transparency problem. “OpenAI has said that the Department of War contractually agreed not to use ChatGPT in agencies that surveil American people,” Carson said. “They have been happy to show contract language when it benefited them, but they refuse to release to the public this contractual provision.”

The full text of OpenAI’s Pentagon contract has not been released publicly. The Intercept noted that while it obtained one contract reflecting OpenAI’s classified network deployment, equivalent documents from Anthropic, Google, and xAI remain outstanding from the government’s document production.

Researchers have long warned that without enforceable guardrails, AI systems deployed by government agencies could enable surveillance at a scale and speed not previously possible — processing vast quantities of digital data to track movement, behavior, and associations across entire populations.

The Broader Architecture: AI Labs as Defense Contractors

The picture that emerges from the contract documents is one in which the leading AI companies in the United States have become deeply integrated partners of the defense establishment — not merely vendors selling software licenses, but active participants in shaping how AI tools will be used to augment military operations.

The DoD’s chief digital and AI officer, Doug Matty, framed the July 2025 announcements in explicitly strategic terms: “The adoption of AI is transforming the Department’s ability to support our warfighters and maintain strategic advantage over our adversaries.” For its part, xAI simultaneously launched “Grok for Government,” a suite of products made available to U.S. government customers through the General Services Administration schedule.

OpenAI had also previously been awarded a separate $200 million contract from the DoD in 2024, and in June 2025 launched “OpenAI for Government” for federal, state, and local government workers. The company also announced a collaboration with defense technology startup Anduril to deploy AI systems for national security missions.

The pattern across all four companies reflects a broader convergence: AI laboratories founded in many cases with stated commitments to safety and human benefit have become significant participants in the defense contracting ecosystem, developing tools explicitly designed for military applications at the highest classification levels.

What the Documents Leave Unanswered

Even with more than 400 pages of contract materials, large portions remain heavily redacted. The specific technical capabilities being developed, the precise military units or commands with access, and the full scope of the information-sharing provisions are not fully visible in the released documents. The Intercept’s FOIA litigation continues, with additional document production expected.

What is visible raises questions that go to the heart of how AI development is governed in the United States: who sets the limits on what these systems can do, how those limits are enforced on classified networks where outside scrutiny is by definition restricted, and what recourse exists when the companies themselves disagree with how their technology is being used.

This article draws on reporting from The Intercept, NBC News, CNBC, OpenAI, and Anthropic.