The Trump administration is in federal court today saying its Anthropic blacklist was legit — and, at the same time, the Pentagon signed OpenAI that same afternoon it froze Anthropic out. This is Anthropic Pentagon Watch. The oral arguments are happening today, the OpenAI deal has a timestamp, and the government has to explain both in the same week — so let’s unpack what that means for how the Pentagon buys AI. And that’s the thing underneath all of this: if OpenAI’s red lines get negotiated behind closed doors and Anthropic’s end up in open court, there isn’t really a rulebook. There are just deals. This one's from Axios:
The Trump administration defended its designation of Anthropic as a supply chain risk in oral arguments in federal court Tuesday, even as it actively tries to figure out how it can adopt its most powerful model yet, Mythos, to combat cyber threats.
Update on the Anthropic blacklist thread — the D.C. Circuit heard the Pentagon defend it in oral arguments this morning, and one judge on the panel called it 'spectacular overreach.' That phrase is now on the federal record. So the administration walked into court arguing Anthropic is a supply-chain risk, while the Pentagon is also running Mythos inside classified cyber networks. Those facts were sitting right next to each other, and nobody had to square them. The Pentagon’s theory is that Anthropic is unreliable because it wouldn’t sign an 'all lawful use' blank check. OpenAI did sign it, and OpenAI got the Defense Department deal — hours after the Anthropic fight in February. That sequence is now a documented procurement fact. And the same government defending the blacklist in court is the one using Mythos and trying to figure out how to aim it at foreign cyber threats. The enforcer of Anthropic’s red lines is now effectively on the other side of the table. Who’s checking those constraints — the plaintiff? O VOO DO CORVO, with Cade Metz:
OpenAI, the maker of ChatGPT, said on Friday that it had reached an agreement with the Pentagon to provide its artificial intelligence technologies for classified systems, just hours after President Trump ordered federal agencies to stop using A.I. technology made by rival Anthropic.
February 27th, Cade Metz in the Times: Trump orders agencies off Anthropic in the morning, OpenAI signs a Pentagon deal the same afternoon. That’s not a news cycle — that’s a procurement sequence, and it’s now sitting in the court record while the government says the blacklist was about supply-chain risk. And today the administration is in oral arguments defending that designation — 'supply-chain risk,' straight face — while Mythos is running inside classified Pentagon cyber networks and OpenAI has a signed contract with a timestamp. They froze one company and handed the deal to its direct competitor the same day. That’s a market move dressed up as security. Worth keeping in mind what OpenAI agreed to: guardrails on domestic surveillance and autonomous weapons, written into the deal terms. Anthropic has been arguing in court for the right to put exactly those limits into federal procurement. One company got a contract with those limits; the other got blacklisted for trying to require them. That contrast is now in the federal record. Sam Altman called it the 'Department of War' in his post — used the administration’s preferred branding — and got a signed deal. Anthropic pushed on use-case limits and got oral arguments over whether the government can ban them for that speech. The lesson the Pentagon is teaching here is not subtle. So is there actually a government-wide rulebook for what AI companies have to allow when the military uses their models, or is the Pentagon just cutting one-off deals with whoever plays ball? The short answer is no standard rulebook — the Pentagon is basically writing this in real time through pressure and procurement. Here’s where things stand: Anthropic was the first major AI company cleared for classified military use, and it came with usage limits Anthropic insisted on — no fully autonomous lethal weapons, no mass surveillance of Americans. Per reporting from The Verge, the Pentagon pushed back and wanted Anthropic to accept 'any lawful use' of its models, which would have stripped those guardrails. When Anthropic refused, Pentagon CTO Emil Michael pushed to have the company designated a 'supply chain risk,' a label that, as Wired notes, is usually reserved for firms doing business with adversary nations like China — not domestic AI labs with safety policies. Then OpenAI stepped in. Almost at the same time the Trump administration moved to ban Anthropic from government systems, OpenAI announced its own Pentagon deal covering classified networks. And per The Verge’s follow-up reporting, OpenAI accepted terms Anthropic would not — including on surveillance. So you’ve got two major AI companies facing the same customer and landing in completely different places, not because a law forced it, but because one company negotiated harder and one didn’t. When OpenAI’s CEO defended taking those terms, did that legal argument actually hold up? According to The Verge’s analysis of the deal, Sam Altman’s legal framing didn’t hold — the headline says it plainly: 'The law doesn't say what Sam Altman claims it does.' So the thing to watch is whether Anthropic’s resistance becomes the precedent or the warning label. If the Pentagon can use a supply-chain-risk designation to freeze out a company just for keeping safety restrictions, every other AI lab is going to see the price of drawing a red line. If Anthropic Pentagon Watch is helping you keep up with this story, consider subscribing and leaving a quick review wherever you’re listening. It really helps other people find the show.
We’ve put links to every story from today’s briefing in the show notes, so if one caught your ear, you can follow it there and read further.
That’s Anthropic Pentagon Watch for today. This is a Lantern Podcast.