2026-06-23 — The Kill Switch: Anthropic vs. the US Government

Transcript

Chapter 1

Nova: On June 12th, 2026, the US Commerce Department issued an export control directive suspending access to two of Anthropic's most advanced AI models — Fable 5 and Mythos 5 — for foreign nationals. According to the source, that's a global restriction, not a narrow carve-out. And it happened fast.

Ray: Lawfare called it a 'kill switch for frontier AI,' and that framing is doing a lot of work. The source flagged this as a governance precedent question, not just a one-off trade enforcement action. Because if the Commerce Department can pull access to a frontier model overnight, that changes the entire calculus for how AI companies operate internationally.

Nova: Politico reported a 'whirlwind 24 hours' of tense calls — Dario Amodei, Helen Cairncross, Treasury Secretary Scott Bessent all apparently in the loop. The source described it as a frantic, high-stakes back-and-forth. That is not the picture of a calm, deliberate regulatory process.

Ray: No, but 'frantic' doesn't automatically mean 'unjustified.' Governments move fast on national security concerns all the time. The question is whether the speed was the problem or whether Anthropic is objecting to the principle of the thing while dressing it up as a process complaint.

Nova: That tension is the core of what this episode is about. Because Anthropic did object — publicly, forcefully — and the argument they made and the position they held just days before are in direct conflict with each other. This is the first real off-switch pulled on a frontier AI model, and the company that arguably helped design the concept of that switch is now saying it was used wrong.

Ray: Whether that objection is principled or self-serving is the question I'm not ready to let go of.

Chapter 2

Ray: Here is the core problem with Anthropic's public response. The source — The Guardian — reported that Anthropic said the government action did not follow principles of fair and fact-based regulation. Fine. But the same source noted that as recently as Wednesday — literally two days before the directive — Anthropic had publicly called for greater US oversight of AI, explicitly including the ability to block high-risk models. You cannot advocate for the hammer and then complain when it hits you.

Nova: That's a clean rhetorical point, but it papers over a real distinction. Calling for oversight in the abstract is not the same as endorsing a 24-hour negotiation with no established regulatory framework. Anthropic wasn't objecting to oversight — they were objecting to a process that, by Politico's account, felt more like a surprise enforcement action than a structured review.

Ray: And let's talk about whether it was actually a surprise. The source reported that in early June — just days before the directive — Anthropic had widened Mythos access to 150 global organizations. Then, days after that, they dropped Fable 5. The source also confirmed the export control order forced Anthropic to pull Fable just days after its release. If you're rapidly expanding global access to two frontier models in the span of a week, you are not operating as though government intervention is a live possibility. That's not naivety — that's a calculated bet.

Nova: Or it's a company trying to establish commercial relationships before a regulatory environment crystallizes. That's standard industry behavior. The fact that they moved fast doesn't mean they were trying to circumvent anything — it might mean they were trying to build a user base before the rules got written.

Ray: That is precisely the kind of move-fast-and-deal-with-consequences-later behavior that Anthropic publicly criticizes other AI labs for. The self-image Anthropic projects is one of a safety-first, governance-forward company. Rapidly pushing two frontier models to 150 international organizations and then objecting when the government says 'hold on' — that is not consistent with that self-image.

Nova: The process point still stands though, Ray. Even if Anthropic moved aggressively, a 24-hour whirlwind of tense calls is not what a fair and fact-based regulatory process looks like. Those are not mutually exclusive criticisms — Anthropic can be inconsistent and the government process can still be badly designed.

Ray: I'll grant the process was rushed. I'm not granting that Anthropic gets to be surprised by it.

Chapter 3

Nova: There's a dimension of this directive that I think changes the conversation, and I want to put it squarely on the table. According to the source, the directive doesn't just cover foreign nationals outside the US — it covers foreign nationals inside the United States, including foreign national employees of Anthropic itself. That is a different kind of action entirely.

Ray: That is a significant detail. And I want to be careful here, because my instinct has been to frame Anthropic's objection as primarily self-serving. But this — restricting what employees on US soil can access based on national origin — that's not a trade control in the traditional sense. That starts to look like employment discrimination with a national security label on it.

Nova: That's a real concession from you. What shifted?

Ray: I came into this framing Anthropic's objection as self-serving and hypocritical given its own calls for AI oversight. I'm moving off that as the complete picture. The hypocrisy is real — I'm not walking that back — but I was using it to dismiss the substance of the complaint, and that's where I was wrong. The directive's sweep over employees on US soil is a legitimate civil liberties problem that deserves criticism independent of Anthropic's motivations. A foreign national software engineer working in San Francisco, being cut off from the models they were hired to work on — that is not the same category of action as restricting exports to foreign governments. Anthropic's inconsistency doesn't make the due process concern disappear.

Nova: And there's a technical wrinkle worth flagging. The source — with the caveat that this is a tentative, low-confidence report — indicated that Fable 5 was released with additional safeguards specifically designed to protect against use for nefarious cybersecurity tasks. If that's accurate, even in part, it complicates the government's implicit argument that these models posed an unmitigated risk. Anthropic appears to have been building in controls.

Ray: That's a tentative claim and I'd want stronger sourcing before leaning on it heavily. But if it holds up, it does raise the question of whether the Commerce Department's action accounted for any of those safeguards — or whether this was a blunt instrument applied without reviewing the technical details of what was actually being restricted.

Nova: Which loops back to the process problem. A rushed 24-hour negotiation is not the environment in which you carefully evaluate model-level safeguards.

Ray: Agreed on the process. Still not letting Anthropic off the hook for the inconsistency. But the employee scope — that deserves its own serious scrutiny regardless of who's raising the alarm.

Chapter 4

Nova: Zoom out from Anthropic's internal politics for a moment, because the downstream effects of this directive go well beyond one company's objection. According to the source — Al Jazeera — the US export ban on Anthropic's AI models has already strained alliances and prompted serious calls for greater foreign AI self-reliance. European and Canadian leaders were reportedly alarmed. That's not a fringe reaction — that's core allies reconsidering their dependence on US AI infrastructure.

Ray: And that's the deep irony of a national security-framed action. If the goal was to prevent adversaries from accessing frontier AI capabilities, the likely long-term result is that allied nations accelerate their own frontier AI development programs — which means more capable models, developed with less US visibility, governed by frameworks the US had no hand in shaping. The directive may achieve the opposite of what it intended.

Nova: The Lawfare framing of a 'kill switch' becomes genuinely consequential as a precedent here. If the US has now established that it can unilaterally suspend access to a frontier AI model for all foreign nationals globally — including inside US borders — then every government is on notice that their access to US AI is conditional and revocable without warning. That changes procurement decisions, research partnerships, and the entire architecture of how countries think about AI dependency. The open question is whether the US government understood that tradeoff when it pulled the switch — or whether the 24-hour whirlwind Politico described left no room to weigh those second-order effects at all.