The White House just lifted the Fable 5 restrictions, and Anthropic's already calling it a safety win — with a number they picked themselves. If you're just catching up: this fight started with Dario Amodei asking Washington for mandatory third-party evals and the authority to block risky frontier releases. Then it got real — government-approved partners reportedly got a frontier Anthropic model while everyone else stayed outside the circle. So the abstract AI-safety pitch turned into a hard question: who gets to run Claude-class systems, and on whose terms? This is Anthropic Pentagon Watch. Today — the ban lifts, Fable 5 goes global, and a 99% figure nobody outside the building has checked. Let's find the missing one percent. From Mike Allen; Sam Sabin at Axios:
The Trump administration lifted export controls on Anthropic's Claude Fable 5 AI model Tuesday evening, with access returning to customers Wednesday, Anthropic said. The move, eagerly awaited by AI developers, restores public access to the company's powerful Mythos-class model that had been pulled for security reasons 18 days ago.
Update on the Anthropic gatekeeper fight — the Commerce Department signed off, and Fable 5 access comes back to customers today. Lutnick says his office 'worked closely with Anthropic to analyze and approve' the model. 'Worked closely with.' The White House pulled it 18 days ago, then put it back. The rules look a lot like a leash, and Lutnick's holding it. Right, and look at the mechanism. This was an export-control lift — a trade-law action. It doesn't touch the procurement fight or the pending litigation. Those are still running separately. And look at the pattern — OpenAI's GPT-5.6 went to a small approved set last week at the government's request, and Mythos 5 got the same treatment. Approval-by-vibes. Nobody's written down what the test even is. Digg writes:
Anthropic is restoring worldwide access to Claude Fable 5 after US export controls were lifted, but only once fresh safety classifiers proved they could stop a reported vulnerability bypass in more than 99 percent of tests; the June suspension had stemmed from real-time nationality verification limits rather than the model's core capabilities.
So the controls come off — we just covered Trump's team lifting them — and Anthropic's headline number is a 99 percent block rate on the vulnerability exploit. Ninety-nine. That's the defendant grading their own test, with no auditor named. And it's a vendor-supplied figure on a fix, which is exactly the kind of number that earns a raised eyebrow before anyone calls it a win. At Pentagon scale, I want to know what the remaining one percent looks like. Here's what gets me — the June suspension was about real-time nationality verification limits. That's an access-control gap, an administrative failure. The classifier patch fixes a different failure mode from the one that actually shut it down. Right, and that split matters. The redeploy leans entirely on Anthropic's own classifier claim — I don't see any third-party verification requirement anywhere in that announcement. Same governance gap, new angle. And it's back globally. Every customer outside the U.S. perimeter who got cut off is back in. Did the classifier update cover the mass-surveillance use case they named as a red line, or was it scoped to the nationality gap? Because those aren't the same thing. And routine coding falls back to Opus 4.8 in the meantime. So the fix ships with false positives they're still tuning — 'over the coming weeks,' per Anthropic. The reassurance is supposed to come from a 'consensus framework' with Amazon, Microsoft, and Google. If this briefing helps you keep up, consider subscribing wherever you’re listening. And if you have a moment, leave a review — it helps other people find the show.
Over the next few weeks, we’re watching whether Anthropic’s classifier refinements reduce false positives for routine coding and debugging requests, and whether the proposed Glasswing consensus framework with Amazon, Microsoft, Google, and other partners moves forward on grading AI jailbreak severity.
We’ve put links to every story from today’s briefing in the show notes, so if one caught your ear, you can dig in there. That’s Anthropic Pentagon Watch for today. This is a Lantern Podcast.