Washington didn't build a new fence this week. It just put a plaque on the one it's been quietly running for months. If you're just joining us: Anthropic's fight over frontier-model safeguards has been running on two tracks — policy pressure for binding rules, and negotiations over who can actually use Claude Fable 5 and Mythos 5. U.S. officials flagged both models as security risks, Anthropic added safety steps, and Commerce later lifted the export-license requirements. Fable's global again, and Mythos access has either been restored or is expanding through the Glasswing program. This is Anthropic Pentagon Watch. Today — a White House deal with the labs, a brand-new Pentagon drone office, and one thing neither of them shipped with. Stick around for that part. We're staying on Anthropic binding AI regulation push — follow the show and you won't miss what comes next. Here's AI Weekly:
The US government is reportedly in advanced talks with OpenAI, Anthropic and Google on a voluntary framework for how the next tier of AI models get released, and the Financial Times reports an announcement could land as soon as next week.
The FT says a formal White House voluntary framework could land as soon as next week — benchmarks, timelines, who gets access at home and abroad. Put simply: OpenAI and Anthropic are already gating frontier releases to government-vetted customer lists. The announcement is the press release. The policy shipped weeks ago. Right. So the emergency model-access fights all spring have quietly hardened into a standing release framework — and the formal version is codifying a process that never had published criteria. And notice the template. Anthropic's Mythos 5 reinstatement was conditioned on cyber defenders and infrastructure providers only. That one negotiated carve-out is reportedly the model for the whole industry framework. One company wrote the first draft of everyone's rulebook. And 'cyber defender' — is that a defined term with an auditable boundary, or a job title you can claim on a customer form? Because OpenAI reportedly told staff the government would approve access customer by customer. In practice, it's an export-licensing queue with no published standard. Michael Scanlon, writing in Federal Times:
Defense Secretary Pete Hegseth has consolidated nearly all of the Pentagon’s drone and autonomous systems programs under a single new office that reports directly to his deputy and will oversee what Chief Pentagon Spokesman Sean Parnell called “the most consequential battlefield innovation of this generation.”
The memo's from June 29 — DRPM-UxS, one office, and it swallows everything that flies, drives, or floats without a human in it. Air, ground, sea, plus the swarming software that runs them. All reporting up to Feinberg. And the timing lines up with the voluntary deal we just walked through. One office consolidates who holds the autonomy leash inside the Pentagon; the White House framework consolidates which models get near it. Supply side and demand side, same week. Parnell calls it 'the most consequential battlefield innovation of this generation.' Great. But the memo never gives you one sentence defining what offensive autonomy actually means, or where the constraint sits. The director slot's still unnamed, so we've got a single joint integrator with no integrator yet. The box is built; the authority's already been pulled from the services, regardless of who fills it. If you follow AI power and accountability, try Musk v Altman Daily — a 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 get podcasts.
What we’re watching next: a formal White House frontier-model framework announcement could arrive as soon as next week.
We’ve put links to every story from today’s briefing in the show notes, so if there’s a story you want to dig into, that’s the place to start.
That’s Anthropic Pentagon Watch for today. This is a Lantern Podcast.