The American kill switch and the $7 billion love note from China

On June 12, at 5:21 in the afternoon, a letter arrived at Anthropic’s offices.

Commerce Department, Bureau of Industry and Security, Howard Lutnick’s signature at the bottom. The message was simple and it was devastating:

“Shut down Fable 5 and Mythos 5 for every user on earth, effective immediately, because some of those users might be foreign nationals.”

And Anthropic did exactly that. Not because they agreed, but because “deemed export” rules don’t give you a choice. Two of the most capable AI systems ever built went dark globally, within hours, pending further review.

Now I want to be fair here: the official reason was a jailbreak. Amazon researchers reportedly found a method to get Fable 5 to surface cybersecurity vulnerabilities in ways that alarmed someone in Washington. Anthropic’s position was that the same capability already exists in OpenAI’s GPT-5.5, which was not touched by the order. A security researcher who reviewed the actual underlying work described it as defensive research, not a weapon; but Anthropic called it “a narrow, non-universal jailbreak” involving asking the model to read a codebase and identify flaws, something their statement pointed out is “used every day by the defenders who keep systems safe.”

But here’s what the mainstream media has not been putting together in the same sentence:

This wasn’t and isolated incident.

In February 2026, the Pentagon wanted Anthropic to waive the contractual restrictions on Claude’s use for two specific things: mass domestic surveillance of American citizens, and fully autonomous lethal weapons, the kind that decide to fire without a human in the loop.

Anthropic refused.

The Trump administration’s response was to label Anthropic a “supply chain risk.” That designation had never before been applied to an American company; it had been used for Huawei, ZTE, actual foreign adversaries.

The government was treating a San Francisco AI lab the way it treats Chinese Communist Party-linked infrastructure companies.

Anthropic filed federal lawsuits in March. Two courts, two different filings. The relationship was not what you’d call cordial.

And then a jailbreak appeared in June, and… well, the models went dark.

Now let’s do one of my favorite intellectual exercises and ask a simple question. If the jailbreak was the real concern, why was GPT-5.5 not touched? Researchers who reviewed the underlying work said Fable’s capabilities in this area were comparable to what GPT-5.5 could already do. The order landed specifically, surgically, on the one company that had spent four months refusing to build the Pentagon’s weapons.

I’m not saying the jailbreak wasn’t real. I’m not saying no security concern existed. What I’m saying is that the pattern of behavior between these two actors, across four months of lawsuits and supply chain designations and Congressional briefings, tells you something that the official explanation alone does not account for. A behavioral read doesn’t end at the stated reason. It starts there and asks what the history of the relationship looks like.

And the answer to that question leads straight to the most ironic plot twist in the entire global AI story of 2026; because on the other side of Washington’s kill switch, something extraordinary happened.

The Little Chinese Lab That Could

A Chinese AI lab called Z.ai announced a new open-source model and its stock jumped more than 30% on the Hong Kong exchange within days. DeepSeek, the Chinese AI startup that has been a recurring nightmare for American AI executives, closed a funding round of approximately $7.4 billion, their largest ever, essentially the same week. On OpenRouter, the platform where developers choose which model to use for their work, the top four most-used models after the ban were all Chinese.

Not one or two. All four.

Washington’s kill switch worked exactly as designed. It just… wasn’t designed for the enemy it was pointed at, I guess.

The vault it was protecting had already been emptied, and months before Howard Lutnick signed that letter. Through a method so mundane, so openly documented, so brazen, that calling it a “black market” almost understates how above ground this operation actually was. And the people who emptied it didn’t need a jailbreak at all. They just needed API keys, fake accounts, and a lot of patience.

Which brings me to the part of this that has nothing to do with jailbreaks or export control lawyers, and everything to do with something most people aren’t tracking yet: