It's got a case number now: Anthropic PBC versus the U.S. Department of War — and yes, that's the defendant's name in the caption. If you're just catching up: Anthropic's Fable 5 and Mythos 5 got briefly cut off after Commerce cited cyber-risk concerns. Anthropic disputed the standard and warned it might freeze frontier releases if that logic spread across the industry. Access came back in stages — but one question was left hanging: whether a government-ordered safety change quietly alters what users actually get. This is Anthropic Pentagon Watch. Today — a docketed lawsuit, a Senate committee vote, and a White House initiative, all landing in the same week. Buckle up. Let's start with that caption. Because the name choice tells you something. If Anthropic binding AI regulation push matters to you, hit follow — we'll be back on it soon. From Ellie Acra at The White House:
The White House, the Department of the Treasury (Treasury), the Department of Homeland Security (DHS) through the Cybersecurity and Infrastructure Security Agency (CISA), and the Department of War (DOW) have worked closely with industry partners to enable faster exploit detection and develop a rapid and prioritized response to cyber vulnerabilities across our critical infrastructure sectors.
The White House gave it a name: GOLD EAGLE, a cybersecurity vulnerability clearinghouse, announced July 14. And look at the sponsor list — Treasury, DHS through CISA, and the Department of War, all coordinating patch response across critical infrastructure. 'Unprecedented cybersecurity vulnerability coordination.' Unprecedented announces a lot and defines very little. What are the intake criteria? Who decides a vulnerability is in scope? Here's the procurement puzzle nobody in that press release wants to touch — CISA already has Mythos embedded. So does Mythos now run its clearances through GOLD EAGLE retroactively, or does the clearinghouse only cover whatever ships next? Right, and CISA is a named partner in the thing that's supposed to coordinate the tool CISA is running. The auditor's inside the audit. Convenient architecture. And the Department of War is on this partner list the same week it's the named defendant in case number 3:26-cv-01996. One arm is building a coordination structure around Anthropic-class tooling; another's in federal court over how it treated Anthropic. Military AI writes:
The US Senate Armed Services Committee has moved proposed military AI guardrails one step closer to becoming law by approving legislation that would regulate how the Pentagon develops, tests, and deploys artificial intelligence and autonomous weapons systems. Included in the committee’s draft of the fiscal year 2027 National Defense Authorization Act, the measure would require meaningful human oversight of AI-enabled weapons, strengthen testing and accountability standards, and prohibit AI from making decisions to launch or detonate nuclear weapons.
Senate Armed Services just cleared military AI guardrails in the FY27 NDAA — meaningful human oversight of AI weapons, and a flat ban on AI making the call to launch or detonate a nuke. A ban. In statute. That's the strongest language I've seen on this beat all year. It's still committee stage, though — this is the draft SASC markup, not a floor vote and not law. The whole NDAA still has to survive the full Senate, the House, and conference. Sure, but look at the timing. Same week Anthropic sues the Department of War for getting punished after asking for guardrails in its contract — Congress starts writing those guardrails into statute so a contractor doesn't have to beg for them clause by clause. And here's my procurement question: the bill wants commanders to retain 'ultimate responsibility' for lethal AI decisions — fine. Does any line of it address vendor concentration? Who supplies the covered system when three companies own the frontier? That's the gap. You can mandate a human in the loop all day, but if the Pentagon depends on one lab's model, the leverage still sits with whoever can flip the access off on a Friday afternoon. Atlantic Council, with Yasmine Abdillahi:
But to institutions around the world, the short-lived order revealed how a foreign government could sever access to frontier AI capability overnight, with no warning and no opportunity for an appeal. The government’s official reason for the suspension was national security concerns. According to Anthropic, the government believed it became aware of a jailbreak technique capable of bypassing Fable 5’s safety controls.
Yasmine Abdillahi's piece gives us the detail we didn't have before — the suspension letter landed Friday, June 12, at 5:21 p.m. Eastern. Friday evening, no advance notice, no restoration timeline. Five twenty-one on a Friday. Somebody wanted a weekend before anyone could pick up a phone and object. And the order swept in Anthropic's own staff, too — any foreign national got cut off from Fable 5 and Mythos 5. That's the detail African institutions read and can't unsee: access severed overnight, no appeal. Here's what gets me. Fable 5 had been public barely a week. Mythos was locked to a tiny Glasswing partner circle. So the 'national security' emergency was over a model almost nobody could depend on yet — and the government still went scorched-earth on it. The Atlantic Council frame is the honest one here: the damage over those three weeks matters less than the signal. The plug exists, and Washington will pull it without warning. Every foreign ministry now writes that into its planning. And the stated reason was a jailbreak on Fable 5's safety controls, which Anthropic called a 'misunderstanding.' Cool. A 'misunderstanding' doesn't usually get your own engineers cut off; it starts to look like leverage dressed up as safety. Here's Civil Rights Litigation Clearinghouse:
This case concerns a challenge to the federal government’s designation of Anthropic PBC as a “supply chain risk” after the company declined to change its usage restrictions. The company alleged that this designation was unlawful and placed significant limitations on its contracting efforts.
There's a case number now. 3:26-cv-01996, Northern District of California, Anthropic versus the U.S. Department of War — and I want to sit in that caption for a second. They named Pete Hegseth personally as a defendant. Filed March 9th. And the theory in the complaint is clean: the Pentagon slapped a 'supply chain risk' designation on Anthropic after the company refused to remove two usage restrictions from its Claude products. Anthropic calls that unlawful retaliation. So the punishment for putting contractual limits on your own product is getting labeled a national-security liability. That's the whole case. Say no to the terms, get blacklisted from the procurement pipeline. And here's the part watchers are chewing on — a judge granted a preliminary injunction in late March, temporarily blocking the ban. But there's a later court notice that's got people nervous about where this actually lands. So the injunction is not the finish line. Ruben Circelli, writing in PCMag:
Before the US government took it offline, Fable 5 was an impressive AI model, with top-notch intelligence outweighing quirks around pricing and more. However, I'm far less enthusiastic about the version that came back after the ban. Fable 5's new restrictions make it clunkier, less performant, and just harder to recommend, especially with new releases from OpenAI (GPT-5.6) and xAI (Grok 4.5) on the scene.
Here's the dollar figure nobody's saying out loud: PCMag's Ruben Circelli says Fable 5 came back clunkier, with more coding restrictions, and Anthropic's rolling back the price hikes because the thing just isn't worth what they wanted to charge. So the government's June 12 letter didn't just block access. It degraded the product, and now the company's eating the pricing consequences. Somebody's absorbing that revenue hit. It's the first confirmation from the consumer press that the capability actually changed — not lab spin, a reviewer with a test bench saying the safety classifier came back significantly more aggressive. That's the tell: access was restored, but the model got worse. That's the user-facing edge on a fight we've mostly watched play out in filings and letters. And it lands the same week they file suit against the Department of War. You break my product, I eat the price rollback, then I see you in the Northern District of California. If you’re tracking Anthropic and the Pentagon, you may also like The Data Center Daily — a daily briefing on AI compute, hyperscaler capex, the power grid, chips, 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, if you want to dig further into anything that stood out. Thanks for listening — that’s Anthropic Pentagon Watch for today. This is a Lantern Podcast.