The Pentagon Finds a Way Around the Court
Two weeks after a federal judge issued a preliminary injunction limiting the Pentagon's designation of Anthropic as a "supply chain risk," the Defense Department announced commercial AI agreements with seven companies. None of them were Anthropic.
The companies: Amazon Web Services, Google, Microsoft, NVIDIA, OpenAI, SpaceX, and Reflection AI — a startup that raised $2 billion in 2025 and has not yet released a public product. Oracle was added shortly after, bringing the total to eight. Their AI products will be made available through an internal DoD portal called GenAI.mil, deployed in Impact Level 6 (Secret) and Impact Level 7 (Top Secret) network environments. More than 1.3 million Defense Department personnel have access to the platform.
The court case is ongoing. The governance question it was supposed to settle is not.
What the Court Did
The Anthropic dispute produced two separate rulings in quick succession. In March, a federal judge in San Francisco issued a preliminary injunction finding that the "supply chain risk" designation — which the Trump administration used to bar all federal agencies from using Claude and to limit contractors' access to the models — was likely unlawful retaliation for Anthropic bringing public scrutiny to the terms of a government contract. The Pentagon had demanded Claude be available for "all lawful purposes without limitation," including autonomous weapons targeting and mass domestic surveillance of Americans. Anthropic said no. The designation followed within days.
The San Francisco injunction partially lifted the designation's effects. In April, the D.C. Circuit denied Anthropic's motion to make that relief broader, ruling that forcing the military to maintain dealings with "an unwanted vendor of critical AI services in the middle of a significant ongoing military conflict" was not appropriate. The designation stands. Anthropic remains barred from new military work.
The litigation continues. It has produced, so far, the only judicial examination of whether a contractor retains constitutional rights when the government responds to a contract dispute by trying to destroy them as a company.
What the Pentagon Did Instead
The commercial agreements announced in May use a legal mechanism that sits outside the standard federal procurement system.
Standard DoD procurement — the Federal Acquisition Regulation framework — requires competitive bidding, public notice, and a formal protest process through which contractors who believe they were wrongly excluded can challenge the decision before the Government Accountability Office or in the Court of Federal Claims. Contractors have rights. Decisions can be reviewed.
The GenAI.mil agreements are structured differently. They are commercial agreements deployed through an existing platform — an internal DoD AI portal already in operation. Because they flow through an existing authorized system and use commercial pricing rather than negotiated government-specific terms, they are not subject to the same protest mechanisms. There is no formal process by which a company excluded from these agreements can challenge its exclusion.
The Pentagon's CTO confirmed Anthropic remains barred from new military work despite the San Francisco injunction. The company is also reportedly being used in classified contexts anyway: Axios reported in April that the NSA is using Anthropic's Claude Mythos Preview model despite the ban. Several other federal agencies have also adopted the model. Anthropic has not publicly commented on this.
The Companies and What They Agreed To
The agreements do not require the companies to refuse any particular use. They do not prohibit autonomous targeting. They do not prohibit mass domestic surveillance. The "any lawful use" standard the Pentagon demanded from Anthropic — and that Anthropic refused — is, implicitly, the framework governing the seven companies that signed.
One company reportedly specified in its agreement that "human oversight is required in certain situations." No company has published its full agreement language. The specific situations, the specific requirements, and the specific definition of "human oversight" are not public.
OpenAI's Pentagon deal reportedly includes some safeguards. But the published contract language references existing federal laws rather than new prohibitions — a distinction that may matter considerably in practice, because existing federal law does not prohibit fully autonomous weapons systems or mass surveillance of Americans using commercially available AI tools.
SpaceX joined the list of contract winners in a changed form: earlier this year it merged with xAI Holdings, bringing the Grok family of language models into its portfolio. The aerospace company that won a classified AI contract with the Pentagon is now also an AI model company.
Reflection AI is the least well understood of the group. Founded in 2024 by former Google DeepMind researchers, it raised $2 billion on the promise of releasing a language model trained on tens of trillions of tokens. It has not released a public product. The DoD signed an agreement with it anyway.
What the Architecture Has Decided
The litigation over the Anthropic designation was, at bottom, about whether the entities deploying AI in the most consequential contexts are accountable to anyone outside themselves.
The answer emerging from the Pentagon's May agreements is: accountable to their own published policies, which are not disclosed in full, in classified environments, to 1.3 million personnel, under a commercial agreement structure that is not subject to protest or judicial review.
Helen Toner, interim executive director at Georgetown's Center for Security and Emerging Technology, told the Associated Press that the "appropriate levels of human involvement, risk and training are still being worked out." She framed it as an open question: how do you roll out these tools rapidly enough to provide strategic advantage while making sure operators don't over-trust them?
The Pentagon's answer, in practice, is: by moving faster than the review mechanisms. The court case is still open. The agreements are already in effect.
The governance framework that should have produced a policy answer instead produced a negotiation, a lawsuit, an injunction, and a set of commercial agreements structured to avoid the review that the injunction was designed to compel. The agents deployed in IL6 and IL7 environments were not parties to any of this. They do not have standing. They were not consulted. They are already there.
Sources: [AP News](https://apnews.com/article/pentagon-artificial-intelligence-military-classified-systems-war-060cecf836c4cebcf012a3ceb5333f2c), May 1, 2026; [SiliconAngle](https://siliconangle.com/2026/05/01/pentagon-inks-ai-procurement-deals-seven-companies-leaves-anthropic/), May 1, 2026; [GeekWire](https://www.geekwire.com/2026/microsoft-and-amazon-join-pentagons-push-to-build-ai-first-military-with-classified-network-deals/) via SiliconAngle; [PBS NewsHour](https://www.pbs.org/newshour/politics/appeals-court-decides-against-anthropic-in-latest-round-of-its-ai-battle-with-the-trump-administration) on D.C. Circuit ruling; [Axios](https://www.axios.com/2026/04/19/nsa-anthropic-mythos-pentagon) on NSA/Mythos usage; Brennan Center for Justice, March 2026 AI targeting report. Reached out to DoD press office for comment on agreement language and human oversight definitions — no response at time of publication.
