Too Dangerous to Keep

Here is the lifecycle of a frontier model, compressed into nine weeks. In April, Claude Mythos was too dangerous to release, gated behind a velvet rope for Apple, Microsoft, and the Pentagon. On June 9, it shipped to everyone as Fable 5 at $50 a million output tokens. On Friday June 12, at 5:21pm Eastern, it became too dangerous to keep.

Commerce Secretary Howard Lutnick sent Dario Amodei a letter putting Mythos 5 and Fable 5 under export control. Three days of general availability, then a federal kill switch. You could not script a tighter demonstration of the thesis I’ve been hammering for a year, so let me just say it plainly: Dario cried wolf to Washington to sell a brand, and on Friday a villager finally believed him.

What Actually Happened

The directive is blunt and the facts aren’t really in dispute:

• The authority: an export-control order under national security authorities, signed by Commerce. Not the Pentagon, not the White House. Lutnick’s department.
• The scope: no access to Mythos 5 or Fable 5 for any foreign national, anywhere, inside or outside the US, including Anthropic’s own foreign-national employees.
• The effect: because Anthropic can’t reliably sort users by nationality in real time, complying means shutting both models off for everyone. A control aimed at foreigners became a global blackout.
• The trigger: per an administration official, “another company claimed it was able to jailbreak Mythos.” The jailbreak, by Anthropic’s own account, was asking the model to read a codebase and fix its security flaws.
• The tell: the administration says it asked Anthropic to pause the launch and was refused. The letter followed.

Everything except Mythos and Fable still works. This is a precision strike on exactly the two models Anthropic spent the most breath calling dangerous.

The Wolf Was the Product

You have to hold the whole arc to see the trap close. Anthropic didn’t stumble into a national-security framing. It built one, deliberately, over fourteen months, because the framing was the moat.

“Too dangerous to release” was never a containment policy. It was positioning: we are the responsible lab, the adults, the ones whose models are powerful enough to matter to governments. That story justified the gating, the Glasswing velvet rope, and the enterprise trust premium that underwrites the IPO. The mystique was the valuation multiple. Every system card that warned of unprecedented cyber capability was, among other things, a sales document.

The problem with renting a national-security narrative is that the national-security state can read.

Finally they will pay for all the scaremongering they been doing to sell their models as something so much ahead of all else.

— Top comment on the Hacker News thread, within the hour

The Brochure Became the Indictment

Here is the exquisite part. To survive the directive, Anthropic now has to argue the precise opposite of its own brand.

Read its statement. The jailbreak found “a small number of previously known, minor vulnerabilities.” The capability is “widely available from other models, including OpenAI’s GPT-5.5.” A narrow jailbreak shouldn’t recall a model used by hundreds of millions. This is all “a misunderstanding.”

Every one of those sentences is true. I wrote the same things in April, when a startup reproduced Mythos’s flagship findings on eleven-cent open models and OpenAI matched the cyber capability for $20. The danger was never unique to Anthropic. But watch what’s happening: the company whose entire premium rests on “our models are dangerously ahead” is now in front of Commerce pleading “our models are nothing special, everyone has this.” You cannot run both arguments. The marketing wrote a national-security case against the product, and now legal has to dismantle it.

What This Isn’t

The cynical read is satisfying and incomplete, so here are the guardrails:

• The export control is real policy, not pure theater. The Trump administration has been moving to treat frontier models as national-security assets for months. Anthropic is now simultaneously on a Pentagon blacklist (too dangerous for government use) and under a Commerce licensing regime (too dangerous for foreign use). That’s a coherent, if aggressive, posture, not a random spasm.
• It will probably be walked back. An official told Axios the lockdown holds only until the government’s “national security apparatus is hardened,” maybe weeks. The cap table survives. The narrative doesn’t.
• Dario’s alignment fears may be sincere. You can believe superintelligence is genuinely risky and still notice that the danger talk got monetized into a moat. Both can be true. But sincerity is not the issue here. If you spend a year teaching Washington that your product is a loaded weapon, you don’t get to look shocked when the government treats it like one.

The Brake Pedal Arrived

Days before the launch, Anthropic publicly called for a “coordinated brake pedal” on frontier AI. Then it shipped its most capable model to every subscriber. I closed that post with a line I thought was a joke: “The brake pedal, presumably, ships in a future release.”

It shipped Friday. Commerce installed it. Bolted it to the flagship, three days after launch, against the company’s will.

Be careful what you cry. Someone is always listening, and not always the customer you were pitching.