The Pentagon quietly broadened how AI helps set targets — and suddenly Anthropic's red lines look like the roadblock somebody wanted gone. If you're just joining, Anthropic's fight with the Pentagon is over whether Claude can do classified military work while the company keeps limits on lethal autonomous weapons and mass surveillance. It has moved way past procurement paperwork: reporting that a version of Claude sits inside the Maven Smart System has raised much sharper questions about targeting workflows, command responsibility, and what happens when a model is asked to support an illegal order. This is Anthropic Pentagon Watch. Today — a revised targeting doctrine, a court split nobody's resolved, and the White House deciding who gets Anthropic's Mythos. Spoiler: not you. Let's start with that Straits Times report on the doctrine, because it changes how the whole blacklisting reads. This one's from The Straits Times:
NEW YORK – The Pentagon has quietly revised its doctrine on how the US military picks its targets in battle, opening the way for artificial intelligence (AI) to make critical wartime decisions in the future. The revised targeting principles, approved without public disclosure in April, envision “systems where AI initiates actions with human monitoring”, evolving from what it describes as the current practice of having “‘human in the loop’ systems, in which a human initiates actions”.
This is the doctrine change that makes the rest of the week make sense. In April, the Pentagon revised its targeting principles — quietly, with no public release — to envision systems where AI initiates actions with human monitoring. That moves away from human-in-the-loop, where a person starts the action. AI initiates, human monitors. They changed who pulls the trigger and announced it to nobody. Bloomberg had to review a document that isn't classified but also isn't public — that's a choice. And it reframes Anthropic's red lines. When they refused certain uses, people could treat it like they were blocking some hypothetical future system. But by April, DoD was already building toward exactly this. So the blacklisting reads as pressure with a purpose: push out the vendor that said no to the thing the doctrine was written to enable. Follow the timeline — April doctrine, then the squeeze. The phrasing in the document is careful, too. 'AI initiates actions with human monitoring.' Monitoring and approving are different things, and they know the difference. Right — 'human monitoring' is the euphemism. At battlefield speed, the person watching the screen is reduced to witnessing the system move. And the doctrine literally cites adversary speed as the reason to take the human off the trigger. Step back for me. If a judge temporarily blocks the Pentagon's Anthropic blacklisting, what changes in the real world? Can agencies and defense contractors go back to using Claude, or can informal national-security warnings still freeze the company out? Yeah, this is where it gets genuinely messy, because two federal courts are pointing in opposite directions. In late March, U.S. District Judge Rita Lin in California granted Anthropic a preliminary injunction. She blocked enforcement of both the governmentwide ban on Claude and the Pentagon's supply-chain-risk designation, writing that the government's actions appeared 'designed to punish Anthropic' rather than protect national security. On paper, that's a real win. But last week, a federal appeals court in Washington, D.C., refused to suspend that same Pentagon blacklisting while the lawsuit plays out. The court said, in its own words, 'the equitable balance here cuts in favor of the government.' So contractors are stuck with conflicting compliance signals: one court says the ban can't be enforced, another says it can stand for now. And in practice, the chill set in before any ruling landed. Reuters reported that major contractors like Lockheed Martin were expected to purge Claude from their supply chains almost immediately after the original ban, because firms weren't going to risk their Pentagon relationships, no matter how the legal fight came out. So even if Anthropic ultimately wins in court, the business damage from contractors self-censoring may already be done? Exactly. Contractors scrambled to distance themselves from Claude even without a formal legal obligation. They're chasing a trillion-dollar annual Pentagon budget and didn't want to land on the wrong side of the administration's preferences. One venture capital firm told CNBC the switch away from Claude 'in no way reflected a perceived shortcoming' of the product — just political risk management. So the thing to watch is whether the California district court's injunction actually gets followed in practice, and whether the D.C. Circuit conflict forces a cleaner answer. Right now, federal contractors still don't know whether using Claude puts their contracts at risk. Singularity.Kiwi writes:
The US frontier AI gatekeeper policy just picked winners and losers — and it’s the incumbents, every time. Mythos is now cleared for release, but the list of approved recipients is opaque, the criteria are unpublished, and the appeals process doesn’t exist. For anyone outside the US security-industrial complex, the message is blunt: the most capable AI is now a controlled export.
The end result of this week's fight is cleaner than I expected: the White House cleared Anthropic to ship Mythos — its frontier cybersecurity model — but only to a government-vetted list of 'trusted partners.' Which means the administration has turned a vendor dispute into a decision about who gets to buy. Legally, that sits closer to export control than procurement — different animal entirely. And look at the structure — opaque list, unpublished criteria, no appeals process. It's a velvet rope, and the bouncer won't even tell you the dress code. New Zealand startups, academics, independent researchers — all on the wrong side by default. The incumbents inside the US security-industrial complex win. Every single time. And this loops back to the Step Back segment we just did: if a court blocks the blacklisting tomorrow, it still doesn't solve this. A 'trusted partners only' channel routes straight around a preliminary injunction. There's no clause to enjoin. Right — they moved the whole fight upstream. You can win your court case and still never get access, because access was never in the contract. If you're following Anthropic’s Pentagon ties, you might also like Musk v Altman Daily — daily court-watch on Elon Musk’s trial against Sam Altman, OpenAI, and Microsoft, covering testimony, exhibits, and the AGI governance fight. Find it wherever you listen to podcasts.
You’ll find links to every story we covered today in the show notes. If one caught your ear, that’s the place to dig in a little further.
That’s Anthropic Pentagon Watch for today. This is a Lantern Podcast.