Few AI stories in 2026 moved as fast, or got as muddled in the retelling, as the fight between the US Department of Defense and Anthropic, and OpenAI’s decision to step into the gap. Within a single week, one AI lab went from holding a classified-network contract to being labelled a national security “supply chain risk”, while its biggest rival announced a deal to run its models on the military’s classified systems.
This post is a timeline built from primary reporting, followed by what each side actually argued and why developers and companies building on AI APIs should pay attention. It isn’t an argument for either side. There are serious people on both.
Background: Anthropic was already working with the Pentagon
The dispute is sometimes described as Anthropic “refusing to work with the military”. That isn’t accurate. According to Mayer Brown’s legal analysis, Anthropic and the Pentagon signed a contract in July 2025 that allowed Claude to be used on classified networks. The disagreement was about the terms of use, not whether to work with defence at all.
The timeline
February 2026: the ultimatum
Defense Secretary Pete Hegseth demanded that AI providers allow their models to be used for “all lawful purposes” without company-imposed restrictions, and set Anthropic a Friday deadline to drop its limits. As NPR reported, Anthropic held to two red lines:
- no use of its models for fully autonomous weapons, on the grounds that current AI isn’t reliable enough for that, and
- no use for mass domestic surveillance of Americans, which the company treats as a fundamental rights issue.
27 February 2026: the ban and the OpenAI announcement
President Trump directed federal agencies to stop using Anthropic’s technology, with a six-month phase-out for some. Hegseth said he would designate Anthropic a supply chain risk, which would bar it from military work and, as he framed it, from commercial dealings with defence contractors.
The same day, OpenAI CEO Sam Altman said OpenAI was working with the Pentagon to deploy its models on classified systems. Altman said OpenAI shared the same red lines, and that its agreement would exclude domestic mass surveillance and autonomous weapons without human approval, while allowing use that complies with legal protections.
That raised an obvious question that critics asked right away: if OpenAI kept similar limits, what exactly had Anthropic refused that OpenAI accepted? The public answer largely came down to framing. OpenAI accepted the “all lawful uses” principle and described its limits as safeguards within it, while Anthropic insisted on contractual restrictions it could enforce.
Early March 2026: the formal designation
The Pentagon formally labelled Anthropic a supply chain risk in early March, a step normally used against companies tied to foreign adversaries. Mayer Brown noted that the administration hadn’t confirmed which legal authority it was relying on. Candidates included the Federal Acquisition Supply Chain Security Act (FASCSA) and 10 U.S.C. ยง 3252. For contractors, the practical effect was new reporting obligations if Anthropic products were used in federal work.
9 March 2026: Anthropic sues
Anthropic filed suit in federal court to reverse the designation, as Axios reported, arguing it was retaliation for the company’s public stance rather than a genuine security finding.
April 2026: an early setback
An appeals court declined Anthropic’s request to temporarily block the designation while the case continued, according to CNBC. That left the label in place for months.
Late August 2026: the district court ruling
In late August, US District Judge Rita Lin in California ruled that the designation was unlawful. TechCrunch’s report says she found it violated the First Amendment as unlawful retaliation and the Fifth Amendment’s due process protections, and that it was arbitrary and capricious. One line from the ruling was widely quoted: “The empty invocation of national security is not a blank check to punish and retaliate against government critics.”
The judge also pointed to contradictions in the government’s position, such as treating Anthropic as a threat while continuing to pursue work with it. A separate Anthropic case in Washington, D.C. was still active at the time of that report, and further appeals were possible, so this story isn’t finished.
The arguments, stated fairly
The Pentagon’s position is that a private vendor shouldn’t be able to veto lawful military operations. Elected officials and the chain of command decide how the military uses tools, and a supplier whose terms of service can cut off capability mid-mission is itself a risk. Supporters also argue that US adversaries won’t hold back, so neither should the US.
Anthropic’s position is that some uses of today’s AI are unsafe or incompatible with civil liberties no matter who authorises them, and that a company has the right, and arguably the duty, to decline to build them. Supporters point out that “lawful” isn’t fixed: what’s legal can change quickly, and a contract that allows “any lawful use” leaves no room to object later.
OpenAI’s position tried to bridge the two: accept government authority over lawful use, keep specific safeguards, and stay at the table. Critics called that a distinction without a difference. Supporters called it the realistic way to have influence over how the technology is used.
Why developers should care
It’s tempting to see this as Washington politics with no bearing on everyday software work. It’s more relevant than that.
- Vendor risk now includes political risk. If your product depends on one model provider, a government decision can suddenly affect whether customers in regulated sectors can use you. Designing for model portability, with an abstraction layer and tested fallbacks, is cheap insurance.
- Acceptable-use policies are real product constraints. Every major AI provider has usage policies. This dispute showed that they’re not boilerplate. Providers may enforce them, and customers may push back. Read them before you build a business on top of an API.
- Government contracts flow down. If you sell to government contractors, their supply chain obligations can become yours. The reporting duties that followed the Anthropic designation applied down the subcontractor chain.
- AI governance is now being decided in courts. Expect more litigation shaping what AI companies can and can’t refuse. The outcome of appeals here will matter well beyond these two companies.
What to watch next
The questions still open as of September 2026 are whether the government appeals the August ruling, how the D.C. case resolves, and whether Congress steps in with rules for military AI use instead of leaving it to contract negotiations. The Center for American Progress, for one, has argued that Congress should settle the red-line questions directly.
Whatever you think of the parties, the episode settled one thing: the terms under which AI models may be used are no longer a footnote in a contract. They’re now one of the main questions in AI policy.
