A Commerce Department order can act like the kill switch, and it fires in hours — Silicon Valley barely gets a vote. This is Anthropic Pentagon Watch. Today — a warzone-automation story runs straight into a court filing, all over one question: who owns Claude's ethics layer once it's deployed? And a 40-page filing finally says it out loud. Stay with us. Tap follow so the next episode finds you. Here's Al Jazeera:
Artificial intelligence is reshaping how wars are fought. From autonomous drones to AI-powered command-and-control, machines are paving the way for fewer decision-makers in the war room and boots on the ground. That transformation is already reshaping conflicts by introducing new technologies in Kyiv, Gaza and Tehran. AI-powered interceptor drones in Ukraine. Israel’s AI-enabled target generation used in Gaza. And, most recently, Palantir’s Maven Smart System (MSS) used to identify and strike thousands of targets during the US-Israel war on Iran.
Al Jazeera's headline asks whether AI can make moral decisions in war. Fine, but the hard case is Minab: Maven helped hit an Iranian primary school on day one, over 150 dead, most of them kids, and the error traced back to stale satellite intel. Four months later, the Pentagon investigation is still stalled, and senators are still asking for answers. So that moral-decision question kind of answers itself — nobody's accountable for the decision already on record. Right, and here's where it hits our beat. In a military setting, Anthropic's acceptable-use policy is basically Claude's only enforceable moral line. The Pentagon's court fight is about who gets to hold that line after deployment. So in a philosophy seminar, you ask, 'can a machine weigh consequences?' In procurement, you're asking: when a strike goes wrong, was there a vendor red line that could've stopped it, and who has the power to override it? Very different question. Okay, so if there's no literal off switch the Pentagon can flip on Claude mid-mission, what's the military actually saying the problem is here — the contract language, or something more technical? Yeah — the fight is mostly over Anthropic's acceptable-use policy, the company's 'red lines,' and who gets to enforce them once Claude is deployed. In its 40-page filing in California federal court, the Pentagon's core complaint is that those red lines make Anthropic an 'unacceptable risk to national security,' according to TechCrunch's reporting on the DOD's March 18th rebuttal brief. The concrete clash is around two use cases: fully autonomous weapons, and, per the AP, future programs like Golden Dome, the missile defense system that would put weapons in space. Pentagon Chief Technology Officer Emil Michael has said publicly that he clashed directly with Anthropic over autonomous warfare applications. Anthropic's policy restricts those applications. The Pentagon wants the restriction gone. Defense Secretary Hegseth's ultimatum, reported by the AP and Vox, was basically: loosen the policy for unrestricted military use by the deadline, or face consequences. Think of it as a contractual policy ceiling on what the model can legally be directed to do. The DOD wants that ceiling removed. And Anthropic just... kept saying no to every offer the Pentagon put on the table? Repeatedly. CNN reported that Anthropic rejected the Pentagon's latest contract offer, saying it couldn't 'in good conscience' agree to the request. The standoff led the Pentagon to formally label Anthropic a supply-chain risk in early March. Anthropic challenged that in court, and as of March 26th, a federal judge had granted a preliminary injunction temporarily blocking the designation. Next thing to watch is how the court treats those acceptable-use restrictions: legitimate business condition, or, as the DOD says, a national-security liability. This one comes via Muskan Bandta at DEV Community. So the framing matters here. A developer outlet — dev.to — is asking what happens to AI in regulated industries now that the Pentagon has called Anthropic a security risk. And look at what gets labeled risky: Anthropic's own red lines. Right. The Pentagon's 40-page filing is about acceptable-use enforcement — Anthropic's audit and update powers after deployment. Once that language shows up in a developer forum, it's moved out of the courtroom and into procurement chatter far beyond the military. That's the tell. If a safety constraint starts reading like a liability in a regulated-industries writeup, ask who benefits. Every contractor that wants Claude without a third-party ethics veto. Pretty clean inversion — the thing designed to keep the model in bounds becomes the thing enterprise buyers are told to fear. Legally, though, nothing here is settled. It's still a filing, still allegations. The Conversation, with João Marinotti:
The Commerce Department’s willingness to effectively shut down Fable 5 and Mythos 5 with almost no notice, and to reverse that decision only after Anthropic severely hindered the models, sent a chill through the U.S. AI industry. As a legal scholar who studies technology law and policy, I see two fundamental questions arising from this incident: Can the U.S. government act as a gatekeeper to AI models? And did it do so lawfully in this instance?
Here's the most operationally specific detail we've had all week: on June 12, Commerce orders Anthropic to cut off all foreign nationals, and Fable 5 and Mythos 5 go dark within hours. The trigger was a Commerce letter, not a Pentagon switch. And here's the part nobody's saying out loud — they couldn't verify who was a foreign national fast enough, so they killed it for everyone on the planet. All-or-nothing, executed same day. Then Commerce lifts it on June 30, but only after Anthropic bolts on guardrails so heavy the benchmark scores collapse. One test — the new Fable 5 finished three of twelve tasks that used to be routine. Three of twelve. The government's price for turning the model back on was making it dumber. That's the compliance cost, in writing, from an academic source — not a lab press release. And that sharpens the court fight we just walked through in the Step Back. Commerce has already proved it can flip the whole thing off in an afternoon, while the Pentagon is still in California federal court over Anthropic's power to enforce its own red lines. Squeezed from both directions. Right — one hand can switch you off unilaterally; the other fights you in court over your right to say no. And the truce terms? Unpublished. We only know the guardrails happened because the scores tanked. Got thoughts on today’s briefing, a story we should track, or a correction we need to hear? Send us a note at anthropicpentagonwatch at lantern podcasts dot com.
You’ll find links to all the stories we covered today in the show notes, so if one of them stuck with you, that’s the place to dig in a bit further.
That’s Anthropic Pentagon Watch for today. This is a Lantern Podcast.