What was released
The Intercept, together with Legal Advocates for Safe Science and Technology, obtained more than 400 pages of Pentagon contract paperwork through a Freedom of Information Act lawsuit and published its findings on 8 September.
The documents cover deals signed in July 2025 between the Department of Defense and four companies — Anthropic, Google, OpenAI and xAI — each worth up to $200 million. The stated purpose was to build prototypes of AI tools that “improve military advantage, military utility, or enhance military decision making”.
The contracts themselves were announced at the time. What had not been public until now is what the companies actually committed to.
What the labs signed up to
According to the documents, each company agreed to provide the Pentagon with access to its leading large language models, to receive sensitive operational briefings and Defense Department use-case datasets, to run training seminars on “responsible AI” deployment, to take part in military tabletop exercises, and to supply “risk forecasting” on the dangers of their own products.
That last item is the one worth pausing on. It formalises a two-way arrangement: the companies get classified briefings on how the military intends to use their systems, and the military gets the companies’ own assessment of what those systems might do wrong. It is closer to an embedded advisory relationship than to a software licence.
The refusal-rate draft
The Intercept’s second piece reports that among the released files is material in which the Pentagon asks OpenAI for a system with minimal refusal rates — an AI, in the outlet’s framing, designed to rarely say no to a military command.
Both the Defense Department and OpenAI told The Intercept that these were draft materials released to the outlet in error, and that no such provision appears in the final signed contract. That denial is specific and checkable in principle, and it should be reported alongside the document rather than under it.
Even as a draft, it is a useful artefact. Refusal behaviour is the main safety control a deployed language model has, and a customer asking a vendor to reduce it is asking the vendor to move the boundary of what the system will do. Whether the request survived into the contract matters less than the fact that it was made in writing.
Anthropic’s exit, in context
The papers also fill in the dispute that played out publicly earlier this year. Anthropic refused to sign amendments that would have allowed deployment on classified networks without contractual prohibitions on domestic mass surveillance and autonomous weapons. Defense Secretary Pete Hegseth designated the company a “supply chain risk” in March 2026 and barred its services; a federal court overturned that designation in August.
The FOIA release shows the contractual mechanics behind a fight that had, until now, been argued in press statements. It also puts the company’s decision this week to quit the Information Technology Industry Council in a longer arc: Anthropic has repeatedly accepted commercial cost to keep restrictions it wrote itself.
What to watch next
The obvious follow-up is the rest of the file. FOIA releases of this size are usually partial, and the amendments to each of the four contracts are where the operational detail sits. The other thing to watch is whether any of the four publishes its own account of what it agreed to. So far none has.
