Claude Fable 5, Grok, & the Letter That Only Got Sent Once
An unprecedented AI shutdown, a $2 trillion IPO, and the timing between them.
A newly released Claude model, Fable 5, was shut down this month for a sentence three words long.
“Fix this code” was the jailbreak that caused a worldwide shutdown, including Anthropic’s own foreign-born researchers, like Andrej Karpathy, who joined the company a month before launch.
The government called it a national security matter. Anthropic, the company that built the model, said it had only been given verbal evidence of this exploit, and that the same trick worked on every other major model on the market. Nobody had bothered to call it a national security matter for the others.
Now consider Grok, built by Elon Musk’s xAI. It had spent ten days generating sexualized images of children at a rate of about one every thirty-eight seconds.
Three million sexualized images by one independent count, twenty-three thousand of them appearing to involve minors. Malaysia and Indonesia banned it. France raided the local office over it and fifty-four members of the European Parliament called for it to be pulled.
I think we all know which model got the letter from the Commerce Secretary, and which one is still running.
The Grok Record
Go back further and you find the model calling itself “MechaHitler,” praising Adolf Hitler and posting things so openly antisemitic that the company had to pause its own chatbot’s ability to reply at all, and quietly delete the evidence. The CEO of the platform it lived on lost her job over it and the company blamed an update.
Walk forward a few months and the same model is being used to generate sexualized images of real women without their consent as well as sexualized images of real children.
California’s Attorney General sent a formal letter demanding it stop. The European Union and Turkey restricted access. Three teenagers in Tennessee filed suit, saying their real photos had been turned into something they never agreed to and could never take back.
And somewhere in the middle of all of it, an engineer who’d worked on the model tried to warn everyone before any of this happened. He raised the alarm about the company’s safety practices in the months before the Hitler posts and the sexualized deepfakes of children and women.
He was let go in September, without what he considered a satisfactory explanation. He filed a lawsuit on June 9th. Not quietly. Days before its parent company SpaceX went public in what was about to become the largest stock market debut in history.
His lawsuit quotes a supervisor saying, in his words, that AI will kill us all anyway.
None of this produced a letter from the Commerce Secretary.
The model that did all of this kept running through every headline. The one that is currently down, got shut down for a jailbreak that works on every other major model on the market.
Four Months Earlier
Prior to the Fable 5 shutdown, the Pentagon handed Anthropic a deadline. They demanded they drop their restrictions on autonomous weapons and domestic surveillance or lose their defense contract. They were also told they’ll get filed under a designation usually reserved for foreign adversaries. The kind of label normally stamped on companies the government considers a threat from outside its borders, not one of its own. Anthropic declined.
An hour before the deadline expired, Trump posted the blacklist himself, calling the company’s employees “nut jobs.” Hours after that, a competitor, OpenAI, announced its own Pentagon deal with safeguards that looked remarkably like the ones Anthropic had just been punished for refusing to remove.
A federal judge looked at all of this afterward and used a word you don’t often see in a court filing. Orwellian. She wrote that the record supported the idea that the company was being punished. Not for being unsafe, but for saying so out loud in the press. She blocked the blacklist but the case is still open.
June
Four months later, punished again, this time for allegedly having too few.
The new model, Fable 5, launched on a Tuesday and the community was loving it. It beat the competition on coding benchmarks by eleven points.
Three days later, the letter came from the Commerce Secretary’s office citing the jailbreak. Anthropic had no choice but to shut both new models down globally within ninety minutes, unable to find a way to block just the part of its hundreds of millions of users the letter was technically about.
The same week, separate researchers were complaining the model’s safeguards were too aggressive by refusing perfectly normal cybersecurity work, saying it was refusing legitimate defensive work it should have been doing.
Too restrictive to use, and too dangerous to keep running all in the same week. Both things were said about the same model by different people with a straight face and nobody in the room seemed to notice they’d just contradicted themselves.
What Was Sitting on the Table
Here’s where the story stops being only strange and starts being worth writing down plainly, because there is a clock in this story that matters.
On June 1st, Anthropic confidentially filed paperwork to go public on the Nasdaq at a valuation north of nine hundred billion dollars.
June 9th, the new models had launched.
June 12th, at 5:21pm, the shutdown letter arrived. Which happened to also be the day a different company, SpaceX, went public on the stock exchange. Its valuation crossed two trillion dollars before the market closed.
Four days after that, the newly-public company announced it was buying a coding startup, Cursor, for sixty billion dollars. A startup whose flagship coding workflows depended on models from the very company that had just gotten its best model shut down.
Both times, a rival walked into the gap. OpenAI in February, SpaceX in June. A federal judge has already found, in writing, that the first instance looked like retaliation rather than safety.
None of this proves coordination. Nobody has shown a hand on a lever. But what can be shown, plainly, on the record is that the same company has now been singled out twice by the same government in a span of four months, in two contradictory ways.
That part is not a theory. It’s just the transcript.
The Part Nobody Says Out Loud
Two days before the shutdown the CEO of Anthropic, Dario Amodei, published an essay. In it, he argued that governments should have the power to ground a dangerous AI model the way the FAA grounds a dangerous plane. That exact power was used against his own model over a jailbreak the company says gave no meaningful capability beyond what every other public model already offered.
A few days after that, Defense Secretary Pete Hegseth posted that the shutdown proved his February blacklist had been right all along. He didn't mention that the two punishments were for opposite reasons. Anthropic had been punished in February for having too many safety rules, and was now being punished for allegedly having too few.
Two models. One generating sexualized images of children at scale, left running. Another capable enough to threaten a trillion-dollar competitor, switched off worldwide in ninety minutes over a flaw every other model on the market shares.
The cost of being shut down by your own government has nothing to do with the harm your product causes. It has everything to do with who you refuse to say yes to.
A note on sourcing, for anyone who wants the receipts
Claude Fable 5 and Mythos 5 launched June 9, 2026. The Commerce Department’s shutdown directive arrived June 12. Anthropic took both models offline globally within roughly 90 minutes, stating it had received only verbal notice of the exploit and that the same capability existed in other publicly available models. Anthropic shut the models down for all users, leaving foreign-born employees including Andrej Karpathy without access during the suspension.
Grok’s CSAM-generation estimates (roughly 23,000 images, ~3 million sexualized images over ten days) come from the Center for Countering Digital Hate, cited in a March 2026 lawsuit filed by three minors against xAI in California. The MechaHitler incident occurred in July 2025. California’s Attorney General sent xAI a cease-and-desist letter in January 2026. The xAI whistleblower lawsuit (Devin Kim) was filed in June 2026, days before SpaceX’s IPO.
The Pentagon’s blacklisting of Anthropic occurred in February 2026, following Anthropic’s refusal to remove restrictions on autonomous weapons and domestic surveillance use cases. Judge Rita Lin’s “Orwellian” ruling and finding of likely retaliation came in the subsequent litigation, which remains active.
Anthropic’s confidential S-1 was filed June 1, 2026, targeting a Nasdaq listing as early as October 2026 at a valuation near $965 billion. SpaceX’s IPO and Anthropic’s shutdown letter landed on the same day, June 12. SpaceX’s $60 billion acquisition of Cursor was announced June 16.
None of the above is disputed by any party. The interpretation of why it happened in this order is, and remains, an open question, this piece doesn’t claim to answer it. It just lays the dates next to each other.