A lab that sued the Pentagon over product controls just spent the week lobbying Congress to build a bigger lead over China. It's the same posture the Pentagon used to justify those controls. Some context before today's development: the White House has accused Chinese startup Moonshot AI of distilling Anthropic's Fable model to build Kimi K3, and of gaining access to restricted Nvidia GB300 systems. That allegation has turned Anthropic's frontier models into a live test for U.S. enforcement. Model access, chip controls, the whole China competition question—it's all riding on one company's tech. This is Anthropic Pentagon Watch. Today—the silence on open weights that Anthropic really doesn't want to talk about, and what the OpenAI deal actually lets the Pentagon do. Buckle up. We're staying with this story: Moonshot Anthropic distillation allegations. Follow the show and you won't miss what comes next. Ana Maria Constantin, over at The Next Web, has the details. So Anthropic's the last major lab that won't sign the open letter defending open-weight models. And per The Next Web, six weeks ago they made that exact argument—for themselves—when Washington killed their access. Right. Anthropic called restrictions on its own powerful model censorship. But with someone else's open weights, suddenly it's no comment. And Silicon Valley spent the whole day pointing at the empty signature line. Nobody wants to say the quiet part out loud—the principle only shows up when the invoice has your name on it. That's the gap Ana Maria Constantin points to—a lab that sued the Pentagon over access limits has no public position on whether open weights should face the same limits. Silence there amounts to a policy. They just won't file it. Convenient policy, too. You get to look principled in court and keep your options open on every competitor's model. Here's The Star:
Washington is moving to strengthen its artificial intelligence posture towards China, with lawmakers advancing new export control bills and the Trump administration deepening cooperation with leading US AI firms as concerns grow that Beijing is rapidly closing the technological gap.
Millions of accounts a week. That's the number Anthropic's national security guy dropped at Aspen—millions of accounts per week they're shutting down for distillation. And we're just taking his word for the scale? Tarun Chhabra puts the lead at six to nine months. He says it would've been a year to eighteen months if not for the Chinese. That's a very precise estimate for something nobody outside the labs can audit. Here's what doesn't sit right. The same company is suing the Pentagon over product controls while its policy head lobbies Congress to extend the lead over China. That's the exact geopolitical goal the Pentagon used to justify the controls Anthropic went to court to kill. Right—persona non grata across government, and still up at the Aspen Security Forum doing free hawk PR. You can't be the plucky First Amendment plaintiff on Monday and the tip of the spear against Beijing on Wednesday. And the distillation fight's only widened since the Moonshot accusations—now it's “virtually all” of China's leading labs, per Chhabra. Convenient framing when you're the one asking Washington to slam the door. Put a dollar sign next to “extend the lead.” Every month in that six-to-nine-month window means enterprise contracts and federal spending. National security and the balance sheet point in exactly the same direction—funny how that works. OpenAI just cut a deal with the Pentagon that Anthropic wouldn't. So what, exactly, did OpenAI agree to? And does it draw the same lines Anthropic was fighting to hold? Short answer: the lines look similar on paper, but critics say the wording is softer where it counts. OpenAI published the agreement on February 28th, then revised it two days later after public backlash. It says its AI systems—quote—“shall not be intentionally used for domestic surveillance of U.S. persons and nationals,” according to OpenAI's own blog post. The updated version also explicitly bars the use of commercially acquired personal data for that surveillance. And the Pentagon affirmed that OpenAI's tools won't be routed through intelligence agencies like the NSA without a separate new agreement. On autonomous weapons, OpenAI has said publicly that the deal protects against that use. But here's the tension: MIT Technology Review reported that Altman himself acknowledged the negotiations were “definitely rushed.” It also reported that OpenAI began pursuing the deal only after the Pentagon publicly reprimanded Anthropic for refusing to drop its restrictions. The following week, the Electronic Frontier Foundation argued that the language was full of what it called “weasel words.” The word “intentionally,” for instance, could give the military significant wiggle room. And a TechPolicy.Press analysis flagged five unresolved issues, including whether the limits extend to foreign surveillance of non-U.S. persons and how the terms would actually be enforced. So that “intentionally” qualifier—is it a meaningful carve-out, or is the EFF reading too much into contract boilerplate? It's still a live legal question; you can't just wave it away as boilerplate. The EFF specifically calls it a loophole that could allow surveillance if it's framed as incidental rather than deliberate. And TechPolicy.Press adds that the deal leaves its enforcement mechanisms undefined. Now we watch whether Anthropic's refusal to accept similar terms leads to any formal supply-chain-risk designation from the Pentagon, and whether Congress or a court ever weighs in on what these private guardrails actually require in practice. This one's from Al Jazeera:
The military AI industry now encompasses everything from cloud computing to autonomous drones and decision-making software. This shift traces back to the US drone programme of the 2000s and 2010s, which turned remote, screen-based killing into a feature of how wars would be fought.
Al Jazeera mapped the military-tech field this week, and one thing jumps out: the injunction Anthropic won sorted the vendor list, but look who's on it. Anduril's Lattice and Palantir are both there. So is the French joint venture Athea, which just took NATO's Enhanced Air Command platform. Nobody paused for the courtroom. And follow the money. On July 10th, the UK signed a 2 billion pound, 15-year deal with a Raytheon-led consortium to train 60,000 soldiers a year on AI. A consortium landing a fifteen-year annuity isn't waiting around to see how Anthropic's suit shakes out. Which makes Anthropic's position even stranger: it's lobbying Congress to extend the U.S. lead while its own competitors work the classified networks unimpeded. They're all chasing the same goal, but Anthropic's on the opposite side of the table. And the whole thing traces back to the drone program of the 2000s—remote, screen-based killing became the template. Now the primes are selling the software layer on top. Palantir's at a defense expo putting up signs. Nobody's hiding. Lattice pulls data from multiple sources into one live picture. Keep your eye on the decision layer here, more than the drone itself. It's already awarded and already running. Have feedback, a story idea, or a correction? Email us at anthropic pentagon watch at lantern podcasts dot com. We’d love to hear from you.
You’ll find links to every story in today’s show notes, so take a look at anything that caught your attention. That’s Anthropic Pentagon Watch for today. This is a Lantern Podcast.