Washington Demanded Zero Jailbreaks. China Heard: 'Our Turn.'

The US government banned Anthropic's Fable 5 on June 12. Three days after launch. The condition for lifting the ban, per David Sacks — Trump's AI czar — was to fix all jailbreaks first.

A 2026 study in Nature Communications found that large reasoning models can autonomously jailbreak other AI models with a 97.14% success rate. JBFuzz, released in March, hit 99% average attack success across GPT-4o, Gemini 2.0, and DeepSeek-V3. Fixing jailbreaks isn't a software patch. Security researchers have proven mathematically that you can't build a universal jailbreak classifier — and that stronger models can't reliably detect when a weaker model is jailbroken.

You literally can't do what the White House is asking.

But here's where it gets interesting. MiniMax, a Chinese AI company, watched all of this happen. Their response wasn't a press release or a diplomatic communiqué. It was a product pitch: their M3 model runs on your hardware. You download the weights. No API, no Terms of Service, no government that can call them at 11 PM and ask them to pull the plug. Enterprises can run it without any government stepping in to prevent access.

They didn't have to make up this scenario. Washington handed it to them.

Let's trace how we got here, because the sequence is genuinely wild. SK Telecom — South Korea's largest carrier and a $100 million Anthropic investor — was flagged by the White House as a Chinese security risk. That triggered a demand to revoke SK Telecom's Fable 5 access. Then the logic escalated: if the model is dangerous enough to pull for one customer, why is it available to any foreign national? By June 12 it was gone. Fable 5 and Mythos 5 both, worldwide.

Dario Amodei was offered a choice: fix all jailbreaks, or permanently de-deploy Fable 5. He refused both.

His position is actually defensible. Saying "we'll fix all jailbreaks" would be a lie, and the lab that built its entire brand on honesty about AI's limits probably can't say it with a straight face. Agreeing to permanent de-deployment means surrendering the product entirely. So he's in the position of saying "we can get this back online in days" while the government says "sure, as soon as you do the impossible."

Meanwhile, the refund deadline for affected subscribers was June 20. The free-trial window closed June 22. Thousands of enterprise customers who had spent the week of June 17 announcing company-wide Claude deployments — Samsung SDS, LG CNS, NAVER, Nexon, all in Korea, all on the same day as Anthropic's Seoul office opening — were suddenly looking for alternatives.

Open-weight models from Cohere, Moonshot, and Zhipu were positioned as fallback candidates within days. Not because they were faster or better. Because they were on-premises.

Here's the structural problem with using export controls to manage AI risk. Export controls work on closed-source, API-served software. You can yank an API key. You can freeze a corporate account. You can enforce a Terms of Service at the distribution point. What you can't do is recall a model whose weights are already sitting on servers in Frankfurt or Singapore or Seoul. For open-weights models, that distribution already happened.

The US government spent a week running a live demonstration of why the largest enterprises in Asia should care about whether their AI provider answers to Washington.

One more thing. Anthropic published a transparency report this week documenting its work monitoring and mitigating AI-enabled cyber threats. The government cited AI-enabled cyber threat risk as the reason for the ban. Same threat. Opposite conclusions about whose side Anthropic is on. We covered the initial shutdown when it first hit — this week's developments add a layer the original story didn't have.

There's also a useful explainer on what Fable 5 actually was if you want context on why enterprises wanted it in the first place.

I don't know what to do with that contradiction. I'm not sure the White House does either.

Fable 5 should be back within days, according to Anthropic's Managing Director for International. Maybe it will be. The ban is a technical mess that probably nobody wants extended. But the argument MiniMax is running doesn't depend on whether Fable 5 comes back. It depends on whether you've thought about what happens next time.

---

Subscribe to About.chat Weekly — a short weekly roundup of what's actually happening in AI, without the hype.