AI
Imagine launching the best product your company has ever made, watching the praise roll in, and then getting a phone call that gives you ninety minutes to take it off the internet. Not in one country. Everywhere. That is roughly what happened to Anthropic in June 2026, and it turned a quiet cybersecurity argument into the biggest AI story of the year. Here is the whole thing, minus the jargon, plus the parts that actually matter to anyone who builds on top of these models.
On June 9, 2026, Anthropic shipped Claude Fable 5 and called it the most capable model it had ever released to the public. It was not marketing fluff. On the company's own launch numbers it scored 80.3% on SWE-bench Pro, well ahead of GPT-5.5 at 58.6% and Gemini 3.1 Pro at 54.2%. One independent newsletter flat out called it the best coding model in the world. It had a one million token context window and a price tag that made it tempting for serious engineering teams.
Three days later, on June 12 at 5:21 p.m. Eastern, Commerce Secretary Howard Lutnick sent CEO Dario Amodei a letter. It ordered Anthropic to block access to Fable 5 and its more powerful sibling, Mythos 5, for any foreign person anywhere on earth. That included Anthropic's own employees who were not US citizens. Since the company had no way to filter users by nationality in real time, it did the only thing it could. It pulled both models for everyone, everywhere. As of late June, they were still dark.
This was the first time any government had reached into a commercial AI product and shut it off on national security grounds. The kill switch everyone theorized about turned out to be real, and it had just been used.
Here is the twist that made this irresistible. According to the Wall Street Journal, the warning that reached Washington came from Amazon. Amazon CEO Andy Jassy reportedly told Treasury Secretary Scott Bessent that his researchers had jailbroken Fable 5 and pulled out information useful for cyberattacks. The government moved shortly after.
Now sit with the relationships involved. Amazon is Anthropic's largest investor, with a total stake reported around 13 billion dollars, and Anthropic has committed to spending over 100 billion dollars on Amazon's cloud over ten years. Amazon also sells its own competing AI models. So the company bankrolling Anthropic's infrastructure is the same one that flagged its newest product to federal officials.
One detail makes it sting more. By one reconstruction, Jassy first tried to call Amodei directly, and Amodei did not pick up. White House officials then nudged Jassy toward Bessent, and he made the case the same day. An Amazon engineer reportedly joked that the company had snitched. Whatever you call it, a cybersecurity disagreement between two partners became a federal shutdown order in a matter of hours.
The two sides do not agree on how scary this actually was, and that gap is the whole ballgame.
Anthropic says the evidence was thin. By its account, the government described a narrow, non universal jailbreak that basically amounted to asking the model to read a codebase and fix software flaws, which surfaced a handful of already known minor bugs that other public models can find too. The company argued that recalling a model used by hundreds of millions of people over that would, applied across the industry, halt every new model release from every frontier lab.
The White House tells a different story. David Sacks, a senior tech adviser to the President, said a trusted partner testing Fable found a real jailbreak of the safety guardrails, that the administration asked Anthropic to fix it or take the model down, and that Amodei refused. Lutnick told reporters he feared the models could end up in the hands of military intelligence users in China or Russia.
Outside experts mostly leaned toward Anthropic. One cybersecurity founder said the model could surface bugs in a few programs but was still a long way from genuinely dangerous territory. Another researcher who read the underlying paper said plainly that it was not a jailbreak. The honest summary is that the severity is disputed and partly based on a single source, so treat both versions as claims rather than settled fact.
This is the part that lawyers cannot stop talking about. Commerce used a provision of the Export Control Reform Act meant for emerging and foundational technologies. By multiple legal analyses, it was the first known time US export controls were used to regulate a specific AI model on national security grounds.
The mechanism was a deemed export, the idea that handing controlled technology to a foreign national inside the US counts as an export to their home country. The catch is that AI models accessed remotely through an API are usually not treated as exports at all. The same regulation had been cited three separate times in the past to explain why remote access is not an export. As one analysis put it, the model weights are not leaving the building. People are simply logging into servers Anthropic runs. That is exactly why a separate bill on remote access security started moving through Congress, because the legal footing here is shaky and largely untested.
The timing could not have been louder. The shutdown landed five days before the G7 summit in France. At a working lunch, Amodei sat across from the President and urged democratic nations not to splinter over AI.
The allies were rattled anyway. France's president warned that if the US can turn off the switch from one day to the next, it puts every company relying on American AI at risk. Canada's prime minister called it a lesson about over reliance and said the real mistake would be to shrug and not diversify. The UK formally asked for an exemption for British citizens and was turned down, with one official calling carve outs to export controls completely illogical. The phrase that started echoing in every capital was AI sovereignty, the idea that you cannot build your economy on a tool a foreign government can switch off without warning.
A few details turned this from a news story into a meme. Just two days before the ban, Amodei published an essay arguing that the government should have FAA style power to block or reverse unsafe frontier models. Forty eight hours later, the government used exactly that kind of power on him.
There is a deeper irony too. Anthropic built its brand on being the safety first lab, and it marketed Mythos as a model too dangerous to fully release. As one researcher put it, if you describe your product as a munition in every press release, eventually a government takes you at your word. The shutdown also landed days after Anthropic confidentially filed for a public offering at a valuation near a trillion dollars, which means regulatory risk stopped being theoretical for future shareholders overnight.
Strip away the drama and one lesson remains. If your product depends on a single hosted frontier model, you now have proof that the model can vanish for reasons that have nothing to do with you. The week of the blackout, open weight coding models from several labs became instant backup options for companies that could not risk waking up to a dead API.
You do not need to panic or rip out your stack. You do need a plan B. Build so you can swap models without rewriting everything, keep an open weight fallback you can actually run, and treat any single provider as a dependency rather than a foundation. The companies that already thought this way barely noticed the outage. Everyone else spent a week scrambling.
The models may come back soon. The precedent is not going anywhere. The question this whole episode forces on every builder is a simple one. Who actually controls the AI you build on?