The removal that outlasted its own deadline
Defense Secretary Pete Hegseth set a six-month clock in February, which ran out in August. The Department's confirmation arrived in October, roughly two months late. The reason it took that long has little to do with bureaucracy. Per the BBC, military analysts had woven Claude deep into Maven Smart System, the Palantir-run platform that organizes drone footage and satellite imagery. They were still using it as recently as last week for intelligence analysis, target identification, and one-page briefs for senior commanders.
Swapping a model out of a system like that is a multi-step engineering job, not a settings toggle. Researchers at Georgetown put it plainly: once an advanced model is inside a complex data engine, replacing it is difficult and slow. That detail deserves attention beyond this story. For all the talk of AI as interchangeable software, the Pentagon needed about eight months to unplug one vendor. These models are infrastructure now, and infrastructure resists removal.
The fight was never about the technology
Hegseth's demand was contractual, not technical. In February he gave Anthropic three days to agree that Claude could be used for "all lawful purposes." Anthropic refused to drop its own prohibitions: no fully autonomous weapons, no mass domestic surveillance. It had already agreed to most other military uses, including through a government-specific Claude build. The Pentagon then did something unusual. In March it designated Anthropic a national security supply chain risk, an authority Reuters noted had never before been used against an American company. It had previously been applied to firms tied to foreign adversaries. The designation barred Anthropic from military contracts and barred defense contractors from touching its models. It followed the collapse of negotiations over a reported $200 million contract to run AI on classified systems.
Anthropic says the fallout cost it billions in lost business. That claim is landing as the company prepares an IPO that could value it near $2 trillion.
Two courts, two readings
The legal status is a split decision. In September, the D.C. Circuit upheld the Pentagon's exclusion 2-1, ruling the department had "ample support" for treating a vendor's disclosed use restrictions as a supply chain risk, and rejecting Anthropic's First Amendment claim. One judge dissented, arguing the reading stretches the statute.
A month earlier, a federal judge in California read the same facts the other way: the designation was unlawful retaliation against Anthropic for speaking out, with no valid national security basis. That ruling covers civilian agencies and non-defense contractors, so the practical result is a patchwork. Claude is barred from the military and permitted across much of the rest of the federal government, depending on which courtroom is closer.
Anthropic is considering asking the full D.C. Circuit to rehear the case, and Supreme Court review remains possible. Nothing here is settled law.
What it adds up to
"Supply chain risk" is doing a lot of work in this story. It is a national-security label applied to a domestic company whose offense was putting limits in its own product and declining to remove them on demand. The label won a contract fight the Pentagon could not win at the negotiating table.
The second half of the story is quieter and more important. About eight months to remove one model from one agency's systems. Every company deploying AI inside its operations should read that number twice. Reversibility is not a feature of these systems. It is a project.
Sources
- [1] Android Headlines — Pentagon Removes Claude from Military Systems (Oct 5, 2026)Read source
- [2] National Technology — US appeals court upholds Pentagon blacklist of Anthropic (Sept 2026)Read source
- [3] Digital Watch Observatory — US appeals court backs Pentagon's label of Anthropic as national-security supply-chain risks (Sept 28, 2026)Read source
- [4] Startup Fortune — Pentagon Stays Off Claude Even as Anthropic Expands Into Other Agencies (Oct 2026)Read source
- [5] The AI Career Lab — Is Claude Banned from Government Use? (Oct 2026)Read source
- [6] Zubiqo — Pentagon Officially Purges Anthropic AI from Military Systems (Oct 5, 2026)Read source