Skip to content
NewsRegulation

Pentagon documents show OpenAI was asked for minimal refusal rates and Anthropic exited the follow-on deal

· by Pondero Newsdesk

The short version

The Intercept obtained more than 400 pages of DoD AI contracts via FOIA suit, revealing a draft clause requesting minimal refusal rates, Anthropic's refusal to sign a deployment agreement, and OpenAI engineers embedded in warfighting units.

Pentagon documents show OpenAI was asked for minimal refusal rates and Anthropic exited the follow-on deal

More than 400 pages of Department of Defense AI contracts, obtained by The Intercept through a FOIA lawsuit, showed that OpenAI, Anthropic, Google, and xAI each signed agreements worth up to $200 million to build AI prototypes for military advantage. A draft clause in OpenAI documents requested AI with "minimal refusal rates." Anthropic refused to sign a follow-on deployment agreement after the Pentagon insisted on language the company said conflicted with its usage policies.

What the documents show

All four agreements were signed in July 2025 and valued at up to $200 million each, per The Intercept's FOIA reporting. The stated purpose was to build AI prototypes to "improve military advantage" across the armed forces. Provisions included two-way information sharing, Pentagon briefings on DoD operational missions, and company participation in "threat ideation exercises."

The "minimal refusal rates" language appeared in draft documents linked to OpenAI's contract. Both the Pentagon and OpenAI said the clause was in draft materials incorrectly included in the FOIA production and did not survive into finalized agreements, per The Intercept.

OpenAI signed a classified network deployment deal on February 27, 2026. An amended version of that agreement embedded OpenAI engineers within military units for what the document called "warfighting support settings." Sam Altman described the deal as "definitely rushed," per The Intercept. Google and xAI reached separate follow-on agreements shortly after the initial contracts.

Anthropic did not. The company refused to sign a deployment agreement without contractual prohibitions on domestic surveillance and autonomous weapons. The Pentagon designated Anthropic a "supply chain risk" in March 2026. A federal judge overturned that designation in August 2026, per The Intercept.

Why it matters

The contracts document a gap between public safety commitments and what frontier labs agreed to with the DoD. For operators in regulated industries (defense contractors, federal agencies, enterprise security teams), the record is instructive: Anthropic drew a public line and paid a commercial price in the form of a supply-chain-risk designation; OpenAI accepted embedded engineers in warfighting units and its CEO acknowledged the pace was rushed.

The "minimal refusal rates" draft language, even if disclaimed by both parties, is now a matter of public record. It shows the DoD negotiated against model refusal as a feature to be minimized, not an operational inconvenience to route around. Any enterprise compliance posture that treats AI refusal as a safety control now has a documented federal counterparty that sought to reduce it.

Anthropic's supply-chain-risk designation, later overturned by a federal court, is the first publicly documented case of a frontier AI lab facing government contracting retaliation for a safety-based refusal. That sets a precedent for how safety-motivated policy positions interact with federal procurement law.

What to watch next

Congressional committees overseeing defense procurement are likely to seek the full unredacted contract set. Whether any of the labs voluntarily publish their DoD agreements or a summary of the safety carveouts they negotiated will determine how much of this becomes a usable compliance reference for enterprise buyers weighing DoD-adjacent AI deployments.

Sources