Sovereign AI and the Mythos Kill Switch

On June 9, 2026, Anthropic unveiled its premier 'Mythos-class' models: the general-purpose Claude Fable 5 and the security-focused Mythos 5. The company described them as its most capable public releases to date, with Fable 5 demonstrating exceptional proficiency in identifying software vulnerabilities.

Three days later, on June 12, the Trump administration issued an export controls directive on national security grounds. The order was absolute: block all foreign nationals from accessing the models. This restriction applied to overseas users, foreign citizens within the United States, and Anthropic's own foreign employees.

Unable to filter users by nationality in real time, Anthropic suspended the models worldwide. This marked the first time Washington imposed export controls directly on an AI model rather than its underlying hardware. The long-debated 'kill switch' had become a reality, reigniting global discussions over sovereign AI.

The Amazon Tip and a Test of Loyalty

The sudden restriction was triggered by a tip from a competitor. Amazon researchers bypassed Fable 5's safety guardrails by altering prompt phrasing, extracting vulnerability details for at least four software products. Amazon CEO Andy Jassy reported these findings to Treasury Secretary Scott Bessent and other administration officials, prompting the White House to classify the model as a direct national security threat.

Anthropic contested the decision. The company argued the government's case rested on 'verbal evidence of a potential narrow, non-universal jailbreak' — a level of standard code analysis that competing models like GPT-5.5 perform regularly. If this threshold were applied industry-wide, Anthropic warned, the development and release of new AI models would effectively freeze.

A deeper dispute lay beneath the surface. Earlier this year, Anthropic declined to permit the US military to use its models for fully autonomous weapons or mass surveillance, leading to its inclusion on a Pentagon blacklist and a 'supply chain risk' designation. Occurring on the eve of its initial public offering, the forced shutdown is widely interpreted as an attempt to bring a defiant partner to heel.

Sovereign AI and a Borrowed Switch

Europe felt the impact of this dependency most acutely. The European Union relies on non-EU providers for over 80% of its technology and 70% of its cloud services, while the United States and China control roughly 90% of global AI computing infrastructure. Overnight, this reliance transformed into an existential vulnerability.

In the United Kingdom, hospitals, enterprises, and researchers immediately lost access to Fable 5. Former Security Minister Tom Tugendhat remarked that national security has become 'about code, not cannons.'

French Prime Minister Sébastien Lecornu announced a transition to a domestic Mistral-based model, declaring that France 'cannot depend on the goodwill of particular partners.' Its intelligence services moved to replace Palantir's tools with a French alternative, and in Asia, Japan warned it risks becoming a colony of foreign AI.

Oxford professor Sandra Wachter described the prevailing sentiment to Fortune: 'This is something we have always known, but never really felt. We have never quite felt what it means to be on the short end of the stick — on the side that has to face the kill switch.'

A $490 Billion Package and a G7 Standoff

Coincidentally, the EU had unveiled its 'European Technological Sovereignty Package' on June 3, just before the shutdown. Built around a Chips Act 2.0 and a Cloud and AI Development Act, the initiative targets roughly €422 billion ($490 billion) over a decade and aims to triple EU data-center capacity within five years. 'We cannot afford to depend on others for the technologies that keep our hospitals running,' stated Commission President Ursula von der Leyen.

At the G7 summit in France, a quieter standoff emerged. US Commerce Secretary Howard Lutnick proposed a 'trusted partners' framework to grant allied nations priority access to frontier US AI models. However, rationing access by nationality provoked a sharp backlash. German MEP Alexandra Geese remarked that the proposal indicated 'the US sees Europe as an enemy, not a friend and ally.'

GLM-5.2 Released the Very Same Day

The most telling development occurred across the Pacific. On the exact day Washington restricted access to Fable 5, Beijing-based Z.ai, formerly known as Zhipu AI, released GLM-5.2. It debuted as an MIT-licensed, open-weight model with no usage restrictions or geographic blocks, allowing anyone to download and run it on their own servers free of charge. It was, in effect, a ready-made sovereign AI alternative.

Its performance was no mere marketing claim. Independent research firm Artificial Analysis ranked GLM-5.2 as the leading open-source model, surpassing even several of Google's Gemini models. On the human-evaluated LMArena coding leaderboard, GLM-5.2 became the most powerful accessible model available after Fable 5's removal under US export controls, trailing only Claude Opus 4.8 by a slim margin on the most demanding coding benchmarks.

GLM-5.2's primary advantage is cost, undercutting GPT-5.5 and Claude Opus by several fold. Coupled with DeepSeek, which recently raised $7.4 billion, the competitive landscape is shifting.

As Fortune observed, China is not running the same race. While American developers chase the frontier with massive capital, China—constrained by chip and capital limits—plays a different game built on open source, cost efficiency, and rapid distribution.

Shaken Dominance, but Is a Wake-up Possible?

The irony is striking. The moment the US restricted access to its most powerful model, the most capable open-source model available outside the country was GLM-5.2, a Chinese one. As Washington tightens its export controls, bargaining power shifts toward open source, tilting the industry's center of gravity toward Beijing and accelerating the global push for sovereign AI.

Yet a sober counterargument exists. Anton Leicht of the Carnegie Endowment for International Peace flatly asserts there is 'no short-term waking up' for dependent nations. Since only US firms construct frontier models and Washington controls the underlying silicon required to train them, even a best-case sovereign megaproject would require at least two years to catch up. Proclaiming AI sovereignty is simple; escaping dependence remains as difficult as ever.

Nevertheless, the Mythos shutdown raises a profound question. The issue is not that US technical supremacy has cracked, but that weaponizing this dominance against allies risks ceding both market share and moral ground to rivals. Although Anthropic engineers and White House officials initiated face-to-face negotiations on June 14 to negotiate a truce, the underlying truth remains exposed: most of the world's AI infrastructure relies on a kill switch controlled by another nation, and that switch is real.