Lawsuit reveals Pentagon contracts with OpenAI, Google, Anthropic and xAI: up to $200 million

While the industry debates how to keep future AGI from slipping out of control, more practical questions are being settled in contracts. The Intercept and Legal Advocates for Safe Science and Technology filed a Freedom of Information Act lawsuit and obtained more than 400 pages of documents on Pentagon deals with four leading AI labs. Below is an account of what they found, with a caveat: the information comes from the publication and its sources, and some documents remain withheld.

What the contracts contain

In July 2025, the US Department of Defense signed agreements with Anthropic, Google, OpenAI and xAI. Each has a ceiling of 200 million dollars. The companies committed to building prototype AI tools for military operations, automated decision-making, logistics and intelligence. According to The Intercept, those prototypes underpin tools already running on the department’s classified networks.

This is not the standard “buy a license and use it” arrangement. All four contracts include an information-sharing provision. The labs give the Pentagon access to their most capable models, advise it on AI adoption plans and discuss current and future use cases. The military may share classified data with the companies to test models on real tasks.

Google has a separate assignment: help develop the AI adoption strategy for the Defense Department and its Chief Digital and Artificial Intelligence Office (CDAO). The company was also tasked with explaining how US competitors use AI. Engineers from all four labs must participate in command-post exercises that simulate real scenarios. The Pentagon expected this to “incrementally improve” the models’ capabilities in intelligence analysis, cybersecurity and autonomous systems.

Where the conflict began

The work expanded beyond the prototypes, and amendments were added to the contracts. The government withheld their contents, citing military and commercial secrecy. Anthropic refused to sign the new version without an explicit ban on domestic surveillance and autonomous weapons. In response, Defense Secretary Pete Hegseth declared the company a supply-chain threat and barred government agencies from using its services. In August, a federal judge overturned that decision.

The Pentagon quickly found replacements in Google, OpenAI and xAI. OpenAI’s contract was approved on February 27, and Sam Altman acknowledged that the deal was “definitely rushed.” Under the amendments, OpenAI engineers may be assigned to organizations that support combat operations, including combatant commands and theater components. The “System Oversight” section is entirely redacted, while the xAI and Google contracts have no such section.

Two documents also state that the Pentagon asked OpenAI to make its models less likely to refuse military requests. The department and the company responded that drafts had been included in the file by mistake and that the signed contract contains no such requirement. An OpenAI spokesperson stresses that the technology must not be used for mass surveillance within the US, to control autonomous weapons or to make automated decisions with severe consequences.

Critics remain unconvinced. Heidi Khlaaf of the AI Now Institute argues that labs should not assess risks themselves in matters of life and death: doing so creates a conflict of interest. Cambridge’s Sophia Goodfriend points out that public statements about restrictions came only after the companies signed contracts that mentioned “autonomous systems.” The Wall Street Journal also reported that Central Command used Anthropic’s technology during strikes on Iran despite Hegseth’s ban. It remains unclear how much control the companies have over the use of their models once those models are inside military systems.

What practitioners should do

Three things are worth considering.

First, model usage policies are not fixed. Until early 2024, OpenAI explicitly prohibited military use, while Google removed restrictions on surveillance and technologies capable of causing harm from its principles in 2025. If your product or corporate policy relies on a provider’s ethical commitments, put those commitments in a contract rather than relying on a press release.

Second, a provider’s political risks become supply-chain risks. The Anthropic episode showed that a company can be declared a threat to government customers, leaving you dependent on court rulings. If you work with the US public sector, plan for the ability to switch providers quickly.

Third, keep in mind that some of the information cannot yet be verified. Many provisions are redacted, key claims come from the publication, and the companies dispute some of them. This is not grounds for drawing conclusions about every lab, but it is grounds for demanding transparency: which restrictions are actually written into the contracts, and who checks compliance?

The main takeaway: the boundary between civilian and military AI products is now defined by deal terms, not slogans. Read them more carefully.