Anthropic and OpenAI just co-signed a letter urging Congress to ban AI-developed bioweapons — a funny look for Anthropic after it spent the front half of this week defending itself in federal court. This is Anthropic Pentagon Watch. Today: a Capitol Hill lobbying letter, Paul Scharre putting 13,000 strike targets on the table, and the labs moving from filing complaints to writing law. And I get to ask the obvious thing — the ink on this bioweapons letter isn't dry, but the planes over Iran already flew. Let's start there, because the strategy shift is the story. Tuesday, they had standing from the injunction. Today, they're co-signing legislation. They're playing offense now. Right, and notice what it concedes. If you're asking Congress for a statute, you're admitting your own contract language and usage policies weren't enough to police the thing. That answers the question we left open Tuesday — who polices the no-bioweapons line once the model's inside government systems. The labs just said: not us, not the contract. Make it law. But look at what the letter covers versus what's actually happening. Scharre lists Iran, Ukraine, Gaza, Venezuela — four live theaters where AI is already in the targeting loop. The letter's about future bio. Not one of those four shows up. It's easier to lobby against a weapon nobody's built yet than the strike planning that already booked 13,000 targets. Exactly. Pick the threat that costs you nothing today. The bio guardrail is cheap because the bio campaign hasn't happened — the strike campaigns have. And that's where Scharre's number actually matters. Wednesday we had Admiral Bradley testifying that humans need confidence AI hits only intended targets — pure abstraction. Now there are 13,000 targets behind that principle. Thirteen thousand. The human-oversight requirement turns from a hearing-room talking point into a body count. Watch the convergence, too — Congress is already drafting statutory bans on AI in nuclear control. Now bio. Two legislative tracks running at once, both narrowing what the Pentagon can buy. And Scharre's whole frame is 'how the Pentagon can manage the risks.' Manage. That's the policy-analyst register — not a hard constraint, a knob you turn. That makes the lineup messier. You had the blacklist arm of the Pentagon and the cautious-uniform arm. Now there's a third voice saying AI warfare's already here and it's manageable. And here's my last suspicion — does cooperating with Congress on bio weaken Anthropic's adversarial posture at the Ninth Circuit? You can't tell a judge your usage limits are sufficient and tell Congress they're not, in the same week. That tension's real, and the Pentagon's appellate lawyers are taking notes. Watch which version of Anthropic shows up in which room. WIRED, with Emily Mullin:
The CEOs of several major artificial intelligence companies are urging members of Congress to adopt new laws that would make it harder for bad actors to develop biological weapons using their technology.
OpenAI and Anthropic just signed a letter pushing Congress to pass laws restricting AI-developed bioweapons. That's the same Anthropic that spent this week in court fighting a Pentagon blacklist — now it's walking into the Capitol asking for a statute. So the week opens with a procurement ban and a Truth Social post, and it closes with CEOs in front of Congress drafting bioweapons law. That's a real change in altitude. And remember Tuesday — Congress is already drafting statutory bans on AI in nuclear weapons control. Now there's a bio track too. Two legislative lanes converging at once. Here's what bugs me. Scharre's piece this week puts AI-assisted strike planning behind more than 13,000 targets in the Iran campaign. The letter asks for future guardrails on bio while the planes already flew on something else entirely. Step back for me: when an AI lab sells access to the Pentagon but writes 'no autonomous weapons' or 'no mass surveillance' into the deal, who actually enforces that once the model is living inside government systems — the company, the contract language, or the Pentagon itself? Honestly? Nobody has a clean answer right now, and that's the crisis. From the reporting, enforcement is still mostly contractual — words on paper. OpenAI's reworked Pentagon deal, announced in late February 2026, says, per the company's own website, that 'the AI system shall not be intentionally used for domestic surveillance of U.S. persons and nationals.' But The Verge reported that the legal landscape Sam Altman pointed to doesn't actually say what he said it does, so those contract guardrails may not have the statutory teeth he suggested. The Cloud Security Alliance put it bluntly: the Anthropic-Pentagon breakdown exposed a 'governance crisis' for commercial AI vendors, because once a model is inside classified government infrastructure, the vendor has very little ability to monitor misuse or cut off access. You can't exactly send an audit team into a classified network. And NPR's reporting has Anthropic saying the Pentagon wanted those safeguards removed, which is why talks collapsed — meaning the Pentagon may see this language as negotiable, not as a floor. So if Anthropic walked away and OpenAI stepped in with a 'compromise,' does that mean OpenAI actually accepted weaker terms — and just dressed them up differently? That's the suspicion TechCrunch and MIT Technology Review raised. Sam Altman himself said the negotiations were 'definitely rushed,' which is not the phrase you want near autonomous-weapons guardrails. So watch whether any of these contract commitments ever get independent verification attached — or whether they stay self-reported by the vendor and self-policed by the same agency that wanted the restrictions gone. Indian Strategic Studies, with Paul Scharre:
The U.S. military struck more than 13,000 targets in the war on Iran, and used artificial intelligence to help plan operations. AI tools were used to synthesize intelligence, help prioritize targets, and build strike packages. The battle space is changing, but the age of AI warfare is already here.
Paul Scharre's piece for Indian Strategic Studies puts a number on it — more than 13,000 targets in the Iran campaign, with AI synthesizing intelligence and building strike packages. The age of AI warfare has already flown missions. And that lands the same week Anthropic and OpenAI co-signed a letter to Congress asking for restrictions on AI-developed bioweapons. So one hand's lobbying for future guardrails while Scharre documents the present-tense body count. Right — and look at Scharre's own list. Iran, Ukraine, Gaza, Venezuela. Four live theaters, no hard constraints on any of them. The bioweapons letter conveniently picks the one threat that hasn't happened yet. Remember Admiral Bradley's testimony Thursday — humans needing confidence that AI hits only intended targets? That was an abstract principle. Thirteen thousand strikes is what the principle costs in practice. And notice Scharre's wording — 'how the Pentagon can manage the risks.' Manage. Not constrain, not prohibit. That's the policy-analyst voice telling you the planes already left, so let's optimize the workflow. If you’re following the defense and governance stakes around Anthropic, you may 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, so if one caught your ear, you can follow it there and read a bit deeper.
That’s Anthropic Pentagon Watch for today. This is a Lantern Podcast.