Analysis
The Intercept published contract documents on September 8 -- obtained through a Freedom of Information Act lawsuit -- that detail the scope of Pentagon partnerships with OpenAI, Anthropic, Google and xAI, revealing obligations well beyond what any of the four labs have described publicly. The reporting is based on contracts originally signed in July 2025, each worth up to $200 million, plus later amendments.
Under the agreements, the four companies committed to building prototype AI tools intended to improve military decision-making, warfighting, logistics and intelligence work -- and, notably, to advise the Pentagon on AI strategy, train military personnel, and forecast the risks of their own technology. That last obligation is unusual: it puts each lab in the position of being paid by the Department of Defense to assess how dangerous its own products could be in a military context, a dynamic that sits uneasily next to the labs' separate, public AI-safety research efforts.
Where the labs diverged
The documents show real friction, not uniform compliance:
- OpenAI -- The Pentagon reportedly asked for a model tuned toward "minimal refusal rates," according to DMNews' review of the documents. OpenAI says that specific phrase never appeared in the final signed contract, though the request itself shows the kind of customization pressure a military customer applies to a general-purpose model.
- Anthropic -- refused to agree to terms allowing unrestricted deployment on classified networks unless its existing restrictions on autonomous weapons and domestic surveillance stayed in place. That refusal led Defense Secretary Pete Hegseth to formally designate Anthropic a "supply chain risk" in February 2026, a decision a federal judge later ruled illegal in an August ruling Pulse covered at the time.
- Google and xAI -- the documents show both agreed to the broader deployment terms with fewer public disputes than Anthropic, though the redactions in the released files limit how much detail is available on their specific carve-outs.
The labs' public commercial pitches have generally emphasized enterprise productivity, coding assistants and consumer products -- Anthropic in particular has built a brand around AI safety research and constitutional AI principles. The Intercept's documents show that brand coexisting with contract language obligating direct involvement in warfighting-relevant tool development, a tension none of the four companies have addressed head-on in public statements since the story published.
This isn't the first sign of friction between AI labs and military customers -- the Anthropic blacklist saga has run for months, through an illegal-ruling decision, an appeal, and a reaffirmation attempt, each covered as it happened. What's new here is documentary evidence of the underlying contract terms themselves, rather than just the public fight over one company's restrictions. It also isn't proof of misconduct: signing a defense contract to build decision-support and logistics tools is different from deploying autonomous weapons, and none of the disclosed documents show a lab building or delivering a weapons system.
Congress has not yet scheduled hearings on the disclosure, and neither Google nor xAI has faced the kind of public pressure that forced Anthropic's blacklist fight into court twice this year -- for now, the documents sit in the public record with no oversight action attached to them.
