In three days, Fable: the most anticipated AI model of the year, went from euphoria to a complete blackout. For the first time, the United States deployed export controls as a tool against a language model.
On June 9, 2026, Anthropic launched Claude Fable 5, the first publicly available version of its Mythos family of models, alongside Mythos 5, which was reserved for selected partner companies. Fable 5 blocks responses in high-risk domains such as cybersecurity, while Mythos 5 operates without some of these safeguards.
The reception was overwhelming. Within 24 hours, demonstrations began pouring in. A developer cloned Minecraft in twenty minutes using a single prompt, including biomes, a day-night cycle, and caves. Another recreated all 151 first-generation Pokémon, complete with statistics and evolutions. A third submitted a McKinsey report and asked the model to produce one of equivalent quality. At the time, no one could have imagined that the system would be shut down just three days later.
The system crashed on Thursday, June 11. According to the American news site The Information and Reuters, cybersecurity researchers at Amazon (Anthropic’s largest investor and the cloud provider hosting its models via AWS) successfully jailbroke the system, bypassing Fable 5’s safeguards and effectively making it behave like Mythos. Specifically, the model provided information that could be used to identify vulnerabilities in at least four software programs and, in theory, could convert scripts into exploits.
That same day, Amazon CEO Andy Jassy contacted Treasury Secretary Scott Bessent and several White House officials directly. According to Politico, he was responding to a request from the government, which had sought his assessment of the model’s security risks.
A crisis task force and three phone calls
On the morning of Friday, June 12, the issue reached the highest levels of government. An emergency task force was convened. Around the table sat Sean Cairncross, National Cybersecurity Director; Scott Bessent, Treasury Secretary; Howard Lutnick, Secretary of Commerce; and Susie Wiles, Chief of Staff. The decision was collective: Anthropic would be asked to voluntarily withdraw the model until the vulnerability was fixed. All that remained was to reach Anthropic CEO Dario Amodei.
That was when the first misunderstandings emerged. The White House claimed Amodei was unreachable, reportedly on a “wellness retreat” The company categorically denied this, stating that its CEO was contacted around noon, made himself available within an hour and fifteen minutes, and that other executives had been offered as alternatives in the meantime.
When the CEO joined the call, he held a series of conversations with senior officials. In his view, the government was overreacting. The vulnerability was not a “universal jailbreak” (a method that bypasses all safeguards), but a narrowly scoped workaround. He argued that the same technique could be used to elicit similar responses from widely used competing models, including OpenAI’s GPT-5.5, none of which were subject to the same restrictions. Applying this standard to Anthropic alone would mean taking a model used by hundreds of millions of people offline over a limited jailbreak. Extending it across the industry, he warned, would effectively freeze the deployment of any frontier model. Faced with the CEO’s hesitation, Scott Bessent warned him that he was making “a bad decision.”
This mistrust, in fact, was not new. It had been building for months. In early 2026, tensions escalated after Anthropic refused to allow Claude to be used to develop lethal autonomous weapons or to enable mass surveillance of the U.S. population, before withdrawing from partnership discussions with the military.
Trump Steps In
Later that same Friday, June 12, Howard Lutnick sent a formal letter to Dario Amodei. The document invoked national security authorities and imposed export controls on Fable 5 and Mythos 5. The decision was endorsed by Donald Trump.
This is where the case takes an unprecedented turn. Export controls, traditionally designed for chips, semiconductors, and physical technologies, were applied for the first time to a language model. Their scope extended beyond the United States: Washington was not only restricting access for its own citizens, but also for all foreign nationals, whether in the U.S. or abroad. The directive even affected Anthropic’s foreign employees, including a significant portion of its research team, who were suddenly denied access to their own product.
However, while an IP address can be filtered, nationality or legal status cannot be reliably verified for every user. Unable to distinguish between users, Anthropic cut off access to all its customers worldwide. “Export controls were a last resort, after we begged them for hours to work with us” a senior White House official told Politico. The company, for its part, issued a strongly worded statement criticizing a regulatory process it said should have been transparent, fair, and grounded in technical evidence.
On Monday, June 15, Howard Lutnick publicly defended the decision, saying officials feared the models could be exploited by users linked to military intelligence services in China, Russia, or other countries deemed high-risk. That same day, Anthropic technical executives met in Washington with officials from the Department of Commerce, including Sean Cairncross, to negotiate a resolution to the crisis. As of Thursday, June 18, Donald Trump had taken up the matter, stating that negotiations with Anthropic were progressing well.
Written in French by Zakaria Choukrallah, edited in English by Amina Kadiri
