A court struck down the Pentagon blacklist, then GenAI.mil launched without Claude

Two events, four days apart, and they point in opposite directions. On August 27 and 28, Judge Rita Lin of the Northern District of California ruled that the Pentagon's designation of Anthropic as a national-security supply-chain risk, and the blacklist that followed, violated the First and Fifth Amendments: retaliation for protected speech, and denial of due process. She ordered the government to rescind the directives. On August 31 the Department of Defense launched GenAI.mil anyway, with ChatGPT Mil and Grok alongside the already deployed Gemini, and Claude absent. Pentagon CTO Emil Michael confirmed that the removal of Anthropic will be complete by the end of September. Winning in court and losing the customer in the same week is the whole story of what an acceptable-use policy costs.

What happened

• The dispute started when Anthropic refused to give the Pentagon unrestricted access to Claude for military use. The two red lines named in reporting from NBC, CNBC and Axios: mass domestic surveillance and autonomous weapons.
• The government responded with the supply-chain-risk designation, which functions as a procurement blacklist.
• Judge Lin found the designation unlawful on two grounds and ordered it withdrawn.
• GenAI.mil runs at IL5, the sensitive-but-unclassified level. OpenAI and xAI adjusted their models and obtained IL5 accreditation; Anthropic did not change its terms.
• Per Fortune, TechCrunch and Gizmodo, 1.7 million of more than 3 million DoD personnel are already onboarded.

Why this matters outside defence

I am not writing about the politics. I am writing about what this establishes for anyone who buys or sells AI under a contract with use restrictions, which is everyone.

The ruling says a model provider can enforce its acceptable-use policy against its largest possible customer without that refusal being a lawful basis for exclusion. That is new. Until now, red lines in a vendor's terms were something you could assume would bend under enough procurement pressure. Now there is a precedent that they can hold, and that the state cannot punish the holding. For a procurement team, the practical consequence is that the restrictions in a vendor's terms are real constraints on your roadmap, not boilerplate.

The court made the red line enforceable; the launch showed that enforceable and commercially survivable are different questions, and the vendor answers only the first.

The launch says the other half. A customer that wants a use case the vendor refuses will find a vendor that agrees, and two did within days. If your architecture depends on a single model provider, the provider's ethics policy is a single point of failure in exactly the same way its pricing or uptime is.

What to do about vendor selection

• Read the acceptable-use policy as a requirements document. List every use case on your roadmap, including the ones two years out, and check each against every vendor's terms. Surveillance-adjacent analytics, automated decisions about individuals, and anything defence-adjacent are where the terms diverge most.
• Decide which direction the risk runs for you. A vendor that refuses use cases is a supply risk if you need those cases, and a governance guarantee if you do not; a regulated customer may prefer a supplier with a demonstrated record of saying no.
• Keep the second provider warm. The frontier model router pattern exists for cost and capability, but it is also the cheapest insurance against a policy conflict. The Pentagon swapped models in days because it already ran Gemini.
• Put your own red lines in your own contracts. If your customers are the ones with restricted use cases, write them into terms explicitly, so the refusal is contractual rather than improvised.
• Version your policy the way you version prompts. When a vendor updates terms, diff them and re-run the roadmap check.

The honest limit

The ruling is a district court decision and the government can appeal; nothing here is settled law. The 1.7 million onboarding figure and the end-of-September timeline come from the Pentagon and from press reports, not from anything independently verifiable. And I have no visibility into what OpenAI and xAI actually changed to obtain IL5 accreditation, so I cannot tell you where their red lines now sit, only that they moved. Which is, for a procurement decision, the most important unknown of all.