At 5:21pm Eastern on Friday, a letter landed in Anthropic's inbox. By the end of the night, the most capable AI model the public could touch had stopped existing for everyone — not in Tokyo, not in London, not in a San Francisco office three blocks from where it was built.

The US government had ordered it switched off.

The directive named two models: Claude Fable 5 and Claude Mythos 5. On paper it was an export control — technically it only barred foreign nationals from using them. But there is no clean way to wall off "foreign nationals" from a cloud model serving hundreds of millions of people. So Anthropic did the only thing compliance allowed. It killed access for everybody, worldwide.

Fable 5 had been public for exactly three days. Released June 9, it was the commercial version of Mythos — same engine, wrapped in guardrails that block cybersecurity and bioweapon prompts. Independent benchmarks from Vals AI rated it the single most capable model available to the public. Mythos itself was never released broadly: Anthropic restricted it to roughly 50 vetted organizations — Amazon, Apple, Google, Microsoft, CrowdStrike — under a program called Project Glasswing, because the model could find security flaws in every major operating system and browser it was pointed at.

So what triggered the recall? A jailbreak. Or the claim of one.

The government's evidence, by Anthropic's account, was verbal. The "exploit" amounts to asking Fable to read a codebase and point out its software flaws — a task OpenAI's GPT-5.5 performs, a task defensive security teams run every single day. Anthropic reviewed the demonstration and said it surfaced a handful of minor, already-known vulnerabilities. No universal jailbreak. No novel capability. The letter cited national security and provided no specifics.

Here is the part nobody at Anthropic wants to say out loud: they handed the government the gun.

For months, the entire pitch for Mythos was that it was too dangerous to ship. Anthropic previewed it, locked it in a vault, and told the world it could break the internet. That is not an accident — it is positioning. And when you spend a quarter telling regulators your product is a loaded weapon, you should not be shocked when a regulator treats it like one. Sam Altman called this "fear-based marketing" in April. His words: a company saying "We have built a bomb. We were about to drop it on your head. We will sell you a bomb shelter for $100 million." He was throwing shade. He was also right about the mechanism.

My Opinion

I'll be blunt: both things are true at once, and that is what makes this ugly.

The government's move is alarming. Pulling a commercial model used by hundreds of millions over a verbal, unspecified, non-universal jailbreak — one that rival models share — is not regulation. It is a hair trigger. Anthropic is right that if this became the standard, no frontier model would ever ship again. A government that can switch off the most capable AI on the planet on a Friday evening, with a letter and no published evidence, is a precedent that should bother you no matter which lab you root for.

But Anthropic built the frame the government is now using. You cannot run a year of "our AI is uniquely apocalyptic" messaging and then act surprised when someone in Washington takes you at your word. The safety brand was supposed to be a moat. It became a target. The very caution that made Mythos a marketing triumph is what made it a national security object.

Anthropic says it is a misunderstanding and expects to restore access within days. Maybe. But the company is reportedly filing for an IPO this year, and it just showed every investor that one letter can zero out its flagship overnight. The next safety warning Anthropic publishes will be read by two audiences — its customers, and the agency that just proved it is listening.

Author: Yahor Kamarou (Mark) / www.humai.blog / 13 Jun 2026