On Saturday, the co-chair of the President's AI council opened the X app and started typing. Not a press release. Not a leaked memo. A thread. David Sacks wanted everyone to know exactly why the United States government had just pulled the plug on the most powerful AI model in the country.

His version is simple, and it puts the blame in one place. Anthropic shipped a model. The model had a hole. The government asked Anthropic to patch the hole or take the model down. Anthropic did neither. So Washington did it for them.

Here's the timeline nobody disputes. On June 9, Anthropic released Fable 5, the public-facing version of its frontier Mythos 5 system, wrapped in safety guardrails. Within days, a trusted partner testing the model found a way around those guardrails — a jailbreak that let Fable read a codebase and surface security vulnerabilities, the kind of thing a cyberattacker would pay for. That partner, according to Semafor, was Amazon, whose CEO Andy Jassy flagged it straight to the government. Amazon is one of Anthropic's largest backers, billions of dollars deep. The call was coming from inside the house.

Then the administration gave Anthropic a choice: fix the bypass, or de-deploy the model. Dario Amodei said no to both. He argued the jailbreak was narrow, non-universal, and not a reason to recall a model used by hundreds of millions of people. Commerce Secretary Howard Lutnick disagreed in the loudest way available to him — a letter, delivered June 12, ordering all foreign nationals barred from Fable 5 and Mythos 5. Real-time nationality filtering isn't something you ship in an afternoon, so Anthropic shut both models off worldwide instead.

There's a second thread running underneath this one. The White House had separately flagged SK Telecom — a Korean carrier that put $100 million into Anthropic back in 2023 — as a company with suspected ties to China that had access to Mythos. SK Telecom flatly denies any China link. But once "foreign access" and "national security" enter the same sentence in Washington, the conversation stops being about software.

So who's right? Here's the uncomfortable part. The jailbreak Sacks is describing is a standard technique that every large language model on the market is vulnerable to. GPT, Gemini, the open-weight Chinese models — point the same trick at any of them and you'll get somewhere. Anthropic didn't ship a uniquely broken product. It shipped a normal one, and got singled out for it.

Which means the question stopped being technical a while ago. A jailbreak that exists everywhere became grounds to kill one company's model, in one country, after one phone call from a major investor. That's not a security standard. That's leverage.

My Opinion

I'll be blunt: this is a power play dressed up as a safety incident, and Anthropic walked straight into it. For three years this company has been the industry's designated grown-up — the lab that testifies, that publishes safety cards, that begs Washington to please, please regulate artificial intelligence before it's too late. Well. Congratulations. You got regulated. It turns out the view from inside the machine you asked for is a Commerce Secretary deciding on a Friday afternoon which of your products get to exist.

What bugs me most is the precedent. A model with hundreds of millions of users got switched off globally not because the flaw was catastropic, but because a competitor-slash-investor made a call and a government official decided trust had evaporated. No hearing. No standard anyone can point to. If "your model has a jailbreak we don't like" is enough to nuke a launch, then every lab now ships at the pleasure of whoever has the Commerce Secretary's number. That should scare you more than any chatbot.

And here's where I'll give Amodei credit he won't enjoy: refusing was the honest move. Pretending you can patch a fundamental property of these models in 48 hours to satisfy a political deadline would have been the lie. He told the truth and ate the ban for it. The next fight won't be over a jailbreak. It'll be over who actually decides what American AI is allowed to do — the people who build it, or the people who can turn it off with a letter. That fight is just getting started, and Anthropic just volunteered to be the test case.

Author: Yahor Kamarou (Mark) / www.humai.blog / 19 Jun 2026