Intercept FOIA Suit Releases 400+ Pages on Pentagon AI Contracts; Refusal-Minimizing Language Disputed

The Intercept published more than 400 pages of Department of Defense AI contracting documents on September 8, 2026, obtained through a FOIA lawsuit brought with Legal Advocates for Safe Science and Technology. The documents cover contracts signed in July 2025 with four companies — Anthropic, Google, OpenAI and xAI — each worth up to $200 million, to build prototypes that “improve military advantage, military utility, or enhance military decision making,” spanning warfighting, automated decision-making, logistics and military intelligence.

The disputed clause. Two of the released documents indicate the Pentagon asked OpenAI to minimize its model’s refusals of military requests. The Department of Defense and OpenAI told The Intercept those were draft materials provided in error and that the final contract contained no such provision. That denial is recorded here as their position; the documents are in the public record and the final contract is not.

The significance does not depend on resolving that dispute. Whether or not the language survived into signature, the documentary record now shows the Pentagon asking for it at some stage — a procurement counterparty requesting that a safety behavior be reduced. A refusal is the mechanism by which a model declines a request its operator should not make. Asking a vendor to make refusals rarer is a specification about what the buyer wants the tool unable to resist.

What this adds to the Anthropic sequence already documented here. Canon records the February 2026 chain: Hegseth threatening an Anthropic blacklist (Feb 24), Anthropic rejecting the Pentagon ultimatum over its red lines (Feb 26), the supply-chain-risk designation (Feb 27), and OpenAI’s deal concluded the same day. The FOIA documents supply what that sequence was about: Anthropic refused to sign follow-up agreements for classified-network deployment without contractual prohibitions against domestic spying and autonomous weapons use. The designation followed that refusal. The Intercept reports the designation was later overturned by a federal judge.

It also adds a detail about the counterparty who did sign: OpenAI finalized its classified-network deployment on February 27, in a process CEO Sam Altman is reported to have described as “definitely rushed.”

Sourcing note. Both sources are The Intercept’s own reporting on documents it obtained; this is a single-outlet release, and no independent party has reviewed the 400-page set. The capture ledger recorded three companies; the documents cover four — Anthropic is in the July 2025 cohort, which is what makes its later refusal a withdrawal rather than an absence. The Altman quote and the judicial reversal of the supply-chain-risk designation are as reported by The Intercept and were not independently verified here; the reversal in particular should be confirmed against the docket before it is relied on.