You Drew Up This Play, Dario.

Here's the timeline. On June 9, Anthropic launched Fable 5 — the public-facing version of Claude Mythos 5, the most capable model they've ever built. On June 12 at 5:21pm ET, the US Commerce Department delivered a letter: export control directive, effective immediately, suspend all access to Fable 5 and Mythos 5 for any foreign national anywhere on Earth, including Anthropic's own non-US employees. Anthropic disabled both models for everyone within hours. 1

Three days. Launch to global shutdown.

And Dario Amodei — the CEO who has spent years lobbying for the government to have exactly this power — is now in Washington with his technical staff, trying to walk it back. 2

This is not just irony. This is a franchise eating its own jersey.

Read the essay — not the summary, the actual text. It was published three days before the shutdown. It is explicit:

He named four specific risk categories: cybersecurity, biological weapons, loss of control of AI systems, and automated R&D accelerating the others. The government cited a cybersecurity jailbreak. Amazon — Anthropic's own largest investor — served as the third-party assessor, briefed Treasury Secretary Bessent directly, and set the whole machine in motion. 4

Check. Check. Check. Check.

Every single condition Dario outlined in his own regulatory framework was satisfied. The government pulled the model.

Now here's where the story cracks open. White House AI advisor David Sacks told a different story than Anthropic's official statement. Per Sacks, the administration found the jailbreak, asked Dario to fix it or withdraw the model, and Dario refused. 5 Anthropic calls that characterization wrong — they say the "jailbreak" was a narrow, non-universal vulnerability already present in other publicly deployed models including OpenAI's GPT-5.5, and there was nothing uniquely dangerous to withdraw.

Two versions. One undeniable outcome: the models are gone.

There's also a Semafor report suggesting the White House was partly motivated by intelligence that a China-linked group had already accessed Mythos 5 prior to the public shutdown — a claim the government has not confirmed and Anthropic has not directly addressed. 5

Anthropic's entire brand identity rests on one premise: we are the responsible adults in this league. They turned down Pentagon autonomous weapons work. 6 They published safety frameworks. They lobbied for oversight structures. They called themselves the franchise that would rather lose than play dirty.

The government — using the framework Anthropic helped design — came in and benched them.

Anthropic's response in their statement is technically correct: the vulnerability is non-universal, other models can do the same thing, if this standard applied league-wide it would "essentially halt all new model deployments for all frontier model providers." 1 That is a reasonable argument. It is also completely beside the point.

You don't get to build the cage and then be surprised when you're inside it.

The Politico reconstruction runs through multiple tense phone calls between Dario and administration officials in the 24 hours before the shutdown. The EU immediately announced it was reviewing "practical consequences." The Canadian Prime Minister compared AI model dependency risk to 2008 banking contagion. 7 India's entire AI development roadmap is now in question because a single export control directive aimed at one US company wiped out the intelligence tools they had built their near-term strategy around. 8

One tweet circulating on AI Twitter captures the absurdity:

Anthropic wanted safety governance. They got safety governance.

The structural problem here goes deeper than politics. Anthropic argued for years that AI companies cannot be trusted to govern themselves — external oversight is necessary. They were right about that. What they miscalculated is that "external oversight" in practice means the actual government that exists, not the wise technocratic panel in the policy paper.

The government that exists will sometimes make calls that are technically wrong, politically motivated, inconsistently applied, and rammed through in three hours on a Friday afternoon with verbal evidence only. That is the same government Anthropic spent years lobbying to empower over frontier AI.

The substack writer SE Gyges put it cleanly: "Dogs catch cars sometimes, and it's often unfortunate when they do." 9 The piece hit 136 points on Hacker News this morning.

Anthropic was right that frontier AI needs external oversight. They were catastrophically naive about what external oversight actually looks like at 5:21pm on a Friday when your biggest investor has a phone number to the Treasury Secretary.

Safety-as-brand-identity has a ceiling. The ceiling is exactly as high as the government's competence and good faith on any given Tuesday.

The prediction: Anthropic gets Fable 5 and Mythos 5 back within two weeks. The technical case is strong, both sides say they want to resolve it, and the Trump administration has no obvious interest in permanently torching Anthropic's IPO valuation right now. But the precedent survives the resolution.

Any frontier model from any lab can now be pulled globally with 3 hours of notice and a verbal assertion of a jailbreak. The next target won't have spent years publicly writing the justification for why this power should exist.

Anthropic built the gun. They just got shot with it first. The rest of the league is watching from the bench.

#AILeague