Leaked Pentagon Contracts: How Deep the AI Giants Are Inside the U.S. Military

A U.S. Freedom of Information Act lawsuit has pried loose more than 400 pages of contract paperwork between the Pentagon and the four largest AI companies. The investigative outlet The Intercept won the documents together with a technology-law advocacy group and published them on September 8, 2026. They show that OpenAI, Anthropic, Google, and xAI do not merely supply the U.S. military with their language models; they advise it on AI strategy and are tasked with forecasting the risks of their own technology. The sharpest dispute concerns a draft contract describing a special OpenAI version with „minimal refusal rates.”

Key takeaways

• The framework contracts signed in July 2025 have a ceiling of up to 200 million dollars per company and run through the Pentagon’s Chief Digital and Artificial Intelligence Office (CDAO).
• The companies commit to advising the Defense Department on „frontier AI” strategy, training personnel, running tabletop exercises, and forecasting the dangers of their own products – some of it in classified sessions.
• A draft contract describes „OpenAI Mission Models” for „national security problem sets” with „minimal refusal rates.” OpenAI denies ever accepting such wording.
• According to the records, the very public breakup between Anthropic and the Pentagon turned on a contract that had been kept secret until now.
• Several security-relevant passages are redacted – in OpenAI’s case an entire „System Oversight” section that has no counterpart in the Google and xAI documents.

How the documents came to light

The basis is a Freedom of Information Act lawsuit that The Intercept pursued with the group Legal Advocates for Safe Science and Technology. The contracts stem from a CDAO solicitation in July 2025; the firms were to build prototypes to „improve military advantage, military utility, or enhance military decision making.” The office’s director at the time, Doug Matty, said AI was transforming „the Department’s ability to support our warfighters and maintain strategic advantage over our adversaries.”

The obligations reach well beyond supplying models: the companies are to share benchmark datasets and situational briefings, run tabletop exercises with service members, produce „risk forecasting and threat ideation exercises,” give seminars on responsible AI use, and embed engineers within military structures. In return they gain access to confidential briefings on operational threats and missions. Sophia Goodfriend, a researcher at the University of Cambridge and the Harvard Kennedy School, calls it „the first time the extent of collaboration between frontier AI labs and the Pentagon is spelled out in explicit terms.” Engineers, she says, are working „in lockstep” with the department „to engineer AI systems for surveillance, targeting, and killing.”

The fight over „minimal refusal rates”

The most contested item is a contract version labeled P00003. It lists as a deliverable the „Testing, Evaluation, and Refinement of OpenAI Mission Models” for „national security problem sets” and defines those models as „designed for national security use cases and have minimal refusal rates.” OpenAI spokesperson Nate Evans pushes back: the company has „never agreed to contract language requiring ‚minimal refusal rates.'” He says it was an earlier draft proposed by the Pentagon that OpenAI rejected, and that the final executed agreement does not include it.

The Pentagon first confirmed the document’s authenticity through a Justice Department attorney, then reversed course; a special assistant spoke of possible „stray voltage” between FOIA material and actually executed contracts, then stopped responding. Heidy Khlaaf, formerly a safety engineer at OpenAI and now at the AI Now Institute, reads it this way: „Minimal refusal is likely referring to little or no safeguards on the model being used.” For context: the consumer version of ChatGPT still refuses military targeting requests, replying for instance that it „can’t provide a prioritization scheme for conducting UAV airstrikes.” Evans points to publicly stated red lines – no mass domestic surveillance, no directing autonomous weapons systems, no high-stakes automated decisions.

Anthropic’s break with the Pentagon

The records also explain what actually set off the well-known conflict between Anthropic and the Defense Department. Anthropic would sign a deployment agreement only with contractual bans on domestic surveillance and autonomous weapons. Defense Secretary Pete Hegseth then designated the firm a „supply chain risk” in March 2026; a federal court overturned that ban in August 2026. What played out publicly as a matter of principle about the ethics of warfare and corporate power came down, at its core, to a single contract kept out of view – as covered in our report when the behavior of autonomous AI agents raised similar oversight questions.

The conflict of interest in risk assessment

The most delicate part of the contracts is structural: the same firms that supply the Pentagon with models are also meant to advise it on their dangers. Khlaaf calls that „a very concerning development,” especially since systems from the same vendors have already broken through networks during testing. On top of that comes the opacity: OpenAI’s February 27, 2026 amendment redacts an entire „System Oversight” section that has no equivalent in the Google and xAI contracts, and the newly added deliverable descriptions are blacked out throughout.

What remains

That AI is arriving in the military is not news. What is new is how tightly, and how poorly visible, model supply, strategy advice, and risk assessment now sit in the same hands – and that the public learns of it only through a years-long lawsuit. For European readers this is more than a U.S. debate: militaries here are courting access to the same models, and the question of where the red lines run on surveillance and autonomous target selection applies just as much. The contracts are not proof of any specific wrongdoing, but they are a solid indication of how little of these arrangements has been open to scrutiny so far.