The US government tried to stop attackers and disarmed defenders. The "jailbreak" that motivated the first export control on a commercial AI model was a defensive workflow: asking the AI to review code and fix flaws. In three days, Mythos 5 and Fable 5 were born and shut down globally — and the question that remains is about sovereignty.
The timeline: 3 days from launch to global shutdown
On June 9, 2026, Anthropic launched Mythos 5 and Fable 5. On June 12, at 17:21, Commerce Secretary Lutnick sent an "Is Informed" letter requiring an export license for all foreigners, in all countries. The deadline to comply: approximately 90 minutes. At 00:50 UTC on June 13, both models were shut down globally. Three days of availability.
We tracked this episode in real time. The speed is what terrifies: a Friday letter cut access in 100+ countries without vote, notice, or ally exemption. There was no debate, no process, no carve-out. It was a kill-switch triggered by unilateral political decision of one country — and nobody outside the US had a voice. The message to every ally running US AI infrastructure was delivered in 90 minutes, not in a treaty.
The legal basis: deemed export stretched to cover inference-as-a-service
This was the first time export controls were applied to a commercial AI model via API. The "deemed export" doctrine was stretched to cover inference-as-a-service. According to Gizmodo, the legal basis is "shaky". According to Lexology, the move is "seismic" — and potentially inconsistent with the BIS's own prior opinions.
The precedent is dangerous. If inference-as-a-service is a "deemed export", then any AI model API can be shut down by unilateral decision of the US government. It does not matter where you host, it does not matter who your vendor is — if the model is American, the kill-switch exists. And the legal basis supporting it is, according to Gizmodo, fragile. The BIS had previously held different opinions on whether inference over an API constitutes a controlled export. That prior position was quietly overridden by a Friday letter.
The dual trigger: SK Telecom and Amazon's defensive "jailbreak"
Two events motivated the letter. First: SK Telecom received access to Mythos via Project Glasswing. The US was alarmed by alleged China ties via SK Group. According to SK Telecom, the denial is direct: China revenue of $1.9M in 2024, 7 employees. The size of the Chinese operation is marginal — but it was enough to trigger the kill-switch.
Second: Amazon researchers identified a guardrail bypass in Fable 5. CEO Andy Jassy escalated personally to the Treasury. Amazon is Anthropic's largest investor ($13B) and the company that reported the "jailbreak". What the government called a jailbreak: asking the model to review code and fix flaws. A legitimate defensive workflow.
Katie Moussouris, according to the open letter published at freefable.org on June 14, was clear: "The prompts worked because they were defensive requests, and that capability cannot be removed without making the model worse at fixing bugs and verifying patches." The jailbreak that motivated the export control was the model's ability to help fix code — the same capability every security team wants in an AI assistant.
The open letter: 126+ signatories and the unanswered argument
The open letter at freefable.org, published on June 14, gathered 126+ signatories. Among them: Katie Moussouris, Bruce Schneier, Alex Stamos, Joe Levy (CEO of Sophos), Mikko Hyppönen, Matthew Green, and Casey Ellis (Bugcrowd). The central argument: Mythos is not "uniquely good" at finding flaws. GPT-5.5, the Chinese Kimi 2.7, and open-source models replicate the capability. Pulling US models only advantages adversaries.
The logic is simple: if the capability exists in non-US models, shutting down the American ones does not remove the capability from the world — it only removes the sovereignty of allies over which model to use. The G7 in Évian ended with no binding agreement. The US blocked multilateral commitments. No official response to the letter's argument was presented.
The sovereignty shock and the structural irony
The UK asked for a carve-out. Refused. According to Macron: "no one will buy AI from the US if they fear it can be turned off at any moment." According to Capgemini, European alternatives cost up to 40% more. The cost of sovereignty is real — but the cost of having no sovereignty is higher. Anyone relying on a US API now knows it can go dark in 90 minutes, with no appeal and no ally exemption.
The structural irony is the detail that matters most. In February 2026, Anthropic removed from its RSP the binding commitment to pause training if safety measures proved inadequate. Now it asks the world to build a verifiable pause mechanism. The same company that weakened its own brakes wants others to install them. It is not minor hypocrisy — it is a signal that the AI safety framework is being built by actors who cannot even maintain their own.
Conclusion: what is your sovereignty plan?
The question for anyone operating AI infrastructure is direct: if a model can be turned off by unilateral political decision of one country, what is your sovereignty plan? We at Tech86 help companies answer exactly that question — auditing dependencies, diversifying jurisdictions, and building contingency plans that do not depend on a single API or a single country. The kill-switch has already been demonstrated. Ignoring it is not strategy — it is a bet.