Megan Skiendel: You know what I keep thinking about? Imagine every major bank agreed to let a federal examiner look at their books before opening day. Every single one. Except one. That one bank isn't breaking a law. But everybody in the room knows their name.
David Sterling: That's Meta right now.
Megan Skiendel: That's exactly Meta right now. OpenAI, Anthropic, Google, xAI, Microsoft — all agreed to submit their models to CASI for federal review. Meta hasn't. And the Trump administration, I mean — they sent emails. Reported by the Times on June 23rd, four people familiar with a confidential request. They didn't exactly whisper.
David Sterling: But it's voluntary. Trump signed the executive order June 2nd — voluntary framework, 30 days to review. So why does being named feel like such a big deal if nobody's forced anything?
Megan Skiendel: Because the word 'voluntary' is doing a lot of work here. Francis Brennan — Meta's spokesman — put out a statement saying Meta 'hopes to sign the agreement soon.' That's not a company planning to refuse. That's a company that got pressure they weren't ready for.
David Sterling: Right. Named publicly as the lone holdout. No law broken. Maximum discomfort.
David Sterling: Well, let's talk about what CASI actually does with a model when they get it. They're looking at capabilities — what can this thing do that could be weaponized, exploited, used to attack infrastructure. Cyberattack potential, adversarial misuse, technology theft. And Muse Spark — Meta launched that in April 2026 — that's the specific model the review request is about.
Megan Skiendel: But okay, what does that look like in practice? Like, has any of this actually surfaced anything real, or is it still theoretical?
David Sterling: David Sterling: Yeah — Anthropic's Mythos model is serious stuff. The Associated Press reported it found vulnerabilities in classified U.S. government systems. That's not a whitepaper scenario. That actually happened.
Megan Skiendel: Wait. During a voluntary review?
David Sterling: During the review process. Which means — the government isn't discovering things companies missed. They're watching companies demonstrate what they already know. CASI has a 30-day window, the review is already operational, and every model that's gone through it has handed the administration a data point Meta doesn't have a counter to.
Megan Skiendel: So Anthropic walks in, Mythos finds a classified-system vulnerability, and suddenly 'voluntary' isn't the frame anymore. The frame is — what does the government now know about your competitors that they don't know about you.
Megan Skiendel: And that's — honestly, that's the Monday morning scenario nobody's writing about. The email from the administration landed Friday. By Monday every other CEO has signed. Tuesday, Meta's board is asking why they haven't. And by the following Friday, Francis Brennan is drafting 'hopes to sign soon.' That's not a principled stand. That's a company that miscalculated the room.
David Sterling: But is this actually about national security? Because the Electronic Frontier Foundation filed an amicus brief — they allege the Pentagon's pressure on Anthropic violated the First Amendment. Retaliation against an uncooperative company, not legitimate security concerns.
Megan Skiendel: Which — wait, who benefits from that framing?
David Sterling: Well, the Trump administration wants minimal regulation — they're running a global AI race argument — and simultaneously wants pre-deployment reviews. Those two things pull in genuinely opposite directions unless the reviews are never actually blocking anyone. Unless the framework is really just — leverage.
Megan Skiendel: Right. And then there's the Llama problem. Meta's model is open-source. I mean — you review it, you restrict it, but it's already deployed everywhere. You can't un-open-source a model. That's not a security review, that's just theater.
David Sterling: Structurally incoherent. That's the actual word for it. But Brennan's 'hopes to sign soon' tells you Meta already knows it. They were never planning a real refusal.
David Sterling: The moment Meta signs — and Brennan's statement makes that close to certain — the Trump administration will have a complete, working pre-deployment review architecture for frontier AI. No vote. No legislation. No public oversight of what CASI actually does with what it finds. Built entirely through emails, public pressure, and one executive order dated June 2nd.
Megan Skiendel: And the voluntary framing — I mean, that was never incidental. That was the mechanism. Every company got to feel like they chose this.
David Sterling: Well. OpenAI, Anthropic, Google, xAI, Microsoft all signed. Infrastructure's operational. Meta's the last gap. Once it closes, the architecture is complete and nobody passed a law.
Megan Skiendel: Yeah. And by the time anyone looked up to ask who approved this, it was already real.
