Right when Anthropic's legal picture looked bleakest, a coalition of establishment conservatives walked into court — for Anthropic. If you're just joining us, Anthropic and the Pentagon have been locked in a procurement fight over Claude on classified military systems. Anthropic says it'll do national-security work, but it draws red lines against lethal autonomous weapons and mass surveillance — while the Pentagon wants access for all lawful purposes. The government designated Anthropic a supply-chain risk, and a D.C. appeals court let that designation stand while the challenge plays out. This is Anthropic Pentagon Watch. Today — who signed that amicus brief, and why the Pentagon's own autonomous-weapons directive might be working against it. Sarah, start with the names. Anthropic-Pentagon supply-chain-risk fight isn't over. Follow us wherever you're listening, and the next chapter comes to you. This one's from Society for the Rule of Law:
The brief argues that the Department of War’s designation of Anthropic as a supply-chain risk does not deserve judicial deference and violates the Constitution. The designation is undeserving of judicial deference because it is a pre-textually motivated retaliation against Anthropic for opposing the Administration’s wishes, because it fails to comply with procedural and statutory standards, and because it subverts national security.
Update on the Anthropic-Pentagon supply-chain-risk fight: rule-of-law amici are now on the record calling the designation pretextual retaliation. It's the first named outside coalition to put its credibility behind Anthropic in the California case. And the name to watch is Alan Raul — Board President of the Society for the Rule of Law. When establishment conservatives walk into court to say the Pentagon overreached, that's a pretty specific signal to sit with. These are serious establishment voices, Sarah — the people who built the national-security legal architecture — and they're filing to say 'sentence first, verdict afterwards.' That quote alone guts the 'trust us, it's classified' defense. The brief says Anthropic doesn't even meet the statutory definition of a supply-chain risk, and the administration made no serious effort to argue otherwise. That's the tell — when you don't even bother to fake the finding, you're admitting it's punishment. Legally, an amicus is just an outside party arguing the public interest — it doesn't bind the court. But it widens the battlefield beyond the two courtrooms where Anthropic was fighting alone. Step back for me: when an AI company says it won't let its model be used for autonomous weapons or mass surveillance, how does that actually get enforced in a Pentagon contract — is there a technical kill switch, third-party audits, financial penalties, or is it basically a handshake? Mostly, it lives in the contract language itself — and that's exactly where this whole dispute cracked open. Anthropic had written restrictions directly into its terms of service, prohibiting Claude from being used for domestic mass surveillance or entirely autonomous weapons targeting. The Pentagon's position, per NPR reporting, is that it doesn't actually intend to use AI in those ways — but it still wants the contractual right to deploy AI 'for all lawful purposes,' full stop, with no vendor-imposed carve-outs. So the 'red line' lives in the agreement as a clause. There's no technical off-switch baked into the model. When Anthropic refused to remove that clause by the Pentagon's deadline, the government didn't just walk away — it designated Anthropic a 'supply chain risk to national security,' a formal procurement blacklisting, per reporting around the early March 2026 action. The Cloud Security Alliance's analysis flags this as the core governance gap: there's no established framework for how commercial AI vendors maintain substantive safety guardrails when those guardrails conflict with what a government customer demands. So if OpenAI then turned around and signed a Pentagon deal — what happened to their stated red lines? OpenAI CEO Sam Altman publicly said his company shares Anthropic's red lines on autonomous weapons and surveillance. But within days of Anthropic's blacklisting, OpenAI announced a deal to deploy AI in classified Pentagon environments — and then, under public and internal pressure, said it would amend that contract over the same surveillance and autonomous-weapons concerns. Here's what to watch: whether 'we'll amend it' produces contract language with teeth, or just softer phrasing that still lets the Pentagon define lawful use. The enforcement mechanism — audits, penalties, technical controls — is still publicly unresolved for every vendor in this space. From CNN:
American artificial intelligence company Anthropic could be at risk being designated a “supply chain risk” — a label typically reserved for companies tied to foreign adversaries. The Pentagon, which uses Anthropic’s Claude AI system on its classified networks, wants broad authority to use it for “all lawful purposes.” But Anthropic has two red lines for the Pentagon: no use in autonomous weapons and no mass surveillance of US citizens.
The Pentagon wants Claude for 'all lawful purposes' — and the lever they're reaching for is a 'supply-chain risk' label. That's the tag you slap on a company tied to a foreign adversary, and they're aiming it at a San Francisco AI firm. Which is the part procurement people choke on. That designation has a specific meaning in the federal supply chain, and 'vendor wrote red lines we don't like' is not what it was built for. And Anthropic's two hard noes are the only specifics in the whole thing: no autonomous weapons, no mass surveillance of Americans. Strip those out and 'all lawful purposes' is just a blank check with Hegseth's name on it. What jumps out, set against the amicus brief we just hit: the rule-of-law coalition isn't defending Claude's politics. They're defending a vendor's right to keep its own contract terms. That's a narrower, sturdier fight than 'AI safety.' The New York Times, with Kevin Roose, Casey Newton:
On Friday, President Trump ordered federal agencies to stop using Anthropic’s A.I. systems and Defense Secretary Pete Hegseth designated the company a “supply chain risk.” Then, just a few hours later, the OpenAI chief executive, Sam Altman, announced that his company reached an agreement with the Pentagon. The deal ensures its technology won’t be used for the same two safety concerns Anthropic raised: domestic mass surveillance or autonomous weapons.
Here's the part that keeps getting soft-pedaled. The NYT Hard Fork piece says OpenAI's deal ensures its tech won't be used for the exact same two things — autonomous weapons and domestic mass surveillance — that Anthropic raised. So the fight was never really about whether the guardrails exist. Both companies got the same two red lines on paper. The difference is who held the pen. That's the procurement read too. Anthropic tried to write its own restrictions into its terms of service, and OpenAI let the Pentagon draft the language. Same protections, different posture toward who controls the contract. And notice the timing on Altman's announcement — hours after Hegseth designated Anthropic a supply-chain risk. Hard to read that as coincidence. A vendor saw the room and signed fast. Reading the room — sure. Or the Pentagon picked the partner who wouldn't argue about the wording, and Anthropic got punished for negotiating in the first place. ESD writes:
Establishes policy and assigns responsibilities for developing and using autonomous and semi-autonomous functions in weapon systems, including armed platforms that are remotely operated or operated by onboard personnel. • Establishes guidelines designed to minimize the probability and consequences of failures in autonomous and semi-autonomous weapon systems that could lead to unintended engagements.
So here's the document everyone's been waving around — DoD Directive 3000.09, signed by Kathleen Hicks in January 2023. The Pentagon's own policy says autonomous weapons are supposed to be designed to allow 'appropriate levels of human judgment' over the use of force. And if THAT'S the standing rule, then turning around and demanding Anthropic strip out guardrails that enforce exactly that — somebody needs to audit that contradiction. Careful, though — the directive establishes policy and an internal working group. It doesn't reach into a vendor's terms of service. As we just walked through, Anthropic's red lines lived in contract language, not in this directive. Which is the interesting tension. The Pentagon has had a written autonomy policy since 2012, updated in '23. So what exactly was Anthropic's restriction duplicating — and who had the authority to waive it? Over on Hacker News:
For what it’s worth, Program Offices and contracting officers are not including this (or any limitations on autonomous control) as a blanket or even specific requirement to be in compliance for any contract awards, so it’s largely irrelevant. There’s also no independent V&V team that will actually verify these kinds of things the way they say they will. DoD used to be ahead in AI now it’s decades behind
And there it is from someone inside the process: contracting officers aren't writing this in as a requirement, and there's no independent verification team. So the directive's a press release with a cover page. If that's accurate, it cuts both ways. The human-judgment language looks great in a brief and enforces nothing on the ground — which undercuts the Pentagon's posture and the 'we already have guardrails' defense at the same time. Hacker News, weighing in:
1.2. POLICY. >a. Autonomous and semi-autonomous weapon systems will be designed to allow >commanders and operators to exercise appropriate levels of human judgment over >the use of >force. Should state: will be designed to mandate and require... "allow" is obviously inserted to permit "wiggle room".
This is the close-read I live for. 'Will be designed to ALLOW' human judgment — not 'mandate,' not 'require.' One verb, and the whole guardrail goes optional. That word choice is deliberate, and the drafters know it. 'Allow' is permissive — it leaves the human in the loop as a design option, not a hard constraint. Hold that next to the rule-of-law brief we opened with. If you follow Anthropic's government work, you might also like The Data Center Daily — a daily briefing on AI compute, hyperscaler capex, the power grid, semiconductor supply, and energy markets reshaped by intelligence at scale. Find it wherever you listen to podcasts.
You'll find links to every story we covered today in the show notes, so if one caught your ear, you can dig into the source material there.
That's Anthropic Pentagon Watch for today. This is a Lantern Podcast.