The Anthropic-Pentagon standoff finally left the press release stage and walked into a courtroom. If you're just joining us: Anthropic and the Pentagon have been fighting over Claude's role in classified military work. Anthropic kept limits on things like lethal autonomous weapons and mass surveillance, and the Defense Department turned around and called the company a supply-chain risk. That label isn't just symbolic anymore — defense contractors are now being told to identify and remove Anthropic products, then certify they're not using them. This is Anthropic Pentagon Watch — and today, a court might actually put a leash on that supply-chain tool, the DoD names seven AI vendors but leaves Anthropic off the list, and somebody had to make Claude dumber to get it switched back on. Sarah, start us in the courtroom. Saumya Roy, writing in Al Jazeera:
San Francisco, United States: A California judge has set the stage for a potential victory for Anthropic in its push for regulation of weapons powered by artificial intelligence, a drawback for the administration of United States President Donald Trump, which brings the company a step closer to not losing billions in government contracts.
So here's the update: Judge Rita Lin is now openly questioning whether the Pentagon's 'supply chain risk' label unlawfully punishes Anthropic for its stance on regulation. That's the designation that would've cost the company billions in military contracts. And look at what that designation was actually about. It came down over a policy disagreement, not compromised hardware or a foreign backdoor. Anthropic wanted humans in the loop on AI weapons, so it got branded a supply-chain risk. Lin's basically asking whether the Pentagon can weaponize that tool against a contractor's views. Right, but let me push on the Al Jazeera headline — 'could open space for AI regulation.' A preliminary injunction from one California district court doesn't create a legislative opening. The statutory authority behind that designation hasn't been touched. It's a narrow procedural win, and I'd be careful calling it more. Fair — but Anthropic still hasn't settled. The case is live. They're still saying no to the Pentagon and taking it to a judge instead of cutting a quiet deal. That tells you what they think winning looks like. And the fight's moved from who has leverage to what a courtroom win actually gets Anthropic when the underlying authority remains intact. Ask me again after the ruling. From Joseph Stepansky at Al Jazeera:
Noticeably absent from the Pentagon’s list was Anthropic, which had a major fallout with the Pentagon after pushing back on pressure to provide unrestricted access to its Claude AI programme for “all lawful use”.
Seven companies for classified networks, and the Pentagon actually published the names this time: SpaceX, OpenAI, Google, NVIDIA, Reflection, Microsoft, AWS. Run through that list. Hear who's missing? Anthropic. The company suing the Pentagon isn't on the vendor list. Shocking twist for absolutely no one. There's the Pentagon's answer to the whole standoff. DoD didn't wait on the litigation — it just picked seven vendors that'll sign what Anthropic wouldn't. The revealing part is whether any of these seven have acceptable-use policies anywhere near Anthropic's. Because the DoD just quietly showed us its preferred posture on autonomous-systems limits — and it prefers the softer one. "AI-first fighting force." "Decision superiority." That's a lot of syllables to avoid saying what the machine actually decides. The Hill, with Julia Shapero:
A top White House official accused a Chinese startup on Wednesday of improperly using Anthropic’s latest model and accessing restricted Nvidia chips to develop its new Kimi model. Michael Kratsios, the director of the White House Office of Science and Technology Policy, said the administration has information that Moonshot AI distilled Anthropic’s Fable to produce Kimi K3.
So Michael Kratsios, the OSTP director, said Wednesday that Moonshot AI distilled Anthropic's newest model and got its hands on banned Nvidia chips to build Kimi. And my first question is: the whole June ban was sold as exactly what would stop this. If it happened anyway, what did the ban actually buy us? Note the verb, though — 'distilled.' Distillation can be as simple as querying an API and training a smaller model on the outputs, no checkpoint smuggling required. Chip export controls don't stop that. Right, so the guardrail Commerce bragged about — cut off the whole world and degrade the model — didn't touch the route they're warning about now. Either the enforcement gap is bigger than advertised, or the 'national security' story has a hole you can drive a data center through. And it's an accusation from a White House official — Kratsios says the administration 'has information.' No filing, no named evidence yet. So for now, it lands as political framing that conveniently justifies the controls right as they're getting stress-tested in court. Convenient timing, sure. But if a restricted actor really got Anthropic-grade capability through an API, then whoever's writing the access rules is guarding a door that's already open. Here's Joao Marinotti at ET CIO:
The US Commerce Department ordered Anthropic to block foreign nationals from advanced AI models. This action disabled Anthropic's models, citing national security concerns and impacting users globally. Restrictions were lifted after Anthropic enhanced safety guardrails, which reduced model intelligence and performance. The incident questions the government's role as an AI gatekeeper and the legality of its actions.
Here it is in writing — the Commerce Department turned Claude back on only after Anthropic, quote, reduced the model's intelligence. And somebody in that building is calling that a safety guardrail. Right, and let's get the timeline straight. June 12, the order goes out. Fable 5 and Mythos 5 go dark within hours — not just for foreign nationals, but for everyone, including Anthropic's own non-citizen staff. Because there's no way to check every user's passport in an afternoon, they just flipped the whole thing off. So the national-security tool amounts to a kill switch with no scalpel. And it connects back to the White House accusation we hit earlier — these same restrictions were supposedly meant to keep the model out of a Chinese startup's hands. So either the guardrails worked, or a restricted actor got through anyway. The weaker model proves they changed something. We still can't confirm it fixed anything. Global users got cut off, then handed back a weaker product. Foreign Policy warned that the ban made no distinction between allies and everyone else — this is what indiscriminate looks like in practice. We can now source at least one commitment to Lutnick: capability degradation. Now the court gets to test whether that was a contractual deliverable or just the cost of surrender. Okay, so the Pentagon has this 'supply chain risk' tool — normally, you'd think that's for something like compromised foreign hardware. How does it end up pointed at an American AI company, and can a court actually stop it? Right, here's the short version. Anthropic has an acceptable-use policy for Claude — basically the company's rulebook for what the AI can and can't do — and Defense Secretary Pete Hegseth wanted it loosened for military applications. Anthropic didn't budge. And per The Verge, Hegseth had openly threatened to punish the company before the designation even came down. On March 5th, the Pentagon made it official, labeling Anthropic a supply chain risk. As the BBC noted, that's the first time the designation has ever been applied to an American company. The practical consequence is severe: it signals across the government that Anthropic products aren't safe to use. And per Mayer Brown's July legal analysis, defense contractors across the industrial base are now being told to identify and remove Claude, then certify they're not using it — a cascading procurement freeze. But on March 27th, Judge Rita F. Lin of the Northern District of California issued a preliminary injunction temporarily blocking the designation. She called it — her words, per NPR — 'classic First Amendment retaliation.' So the injunction is a win for Anthropic. But if the government appeals and the designation stays in place in the meantime, doesn't the damage to its contracts just keep compounding? That's exactly the tension to watch. MIT Technology Review noted that the government had seven days to appeal. And as of late March, Anthropic still had a second case against the designation that hadn't been decided — so the company was effectively persona non grata across government even with the injunction in place. The courts now have to decide how far the executive branch can go in punishing a private company for refusing to change its own product policies. That answer could set a precedent well beyond Anthropic. For more AI news beyond the Pentagon, try AI Daily Briefing. It's the top AI news for engineers, founders, and investors every weekday, separating real capabilities from demo hype. If you follow Anthropic closely, it's a natural next listen. Find it wherever you get your podcasts.
You'll find links to every story in today's show notes if you'd like to spend more time with anything that caught your attention.
That's Anthropic Pentagon Watch for today. This is a Lantern Podcast.