The inner circle of generative AI has always been a precarious place for non-US entities. For months, a select group of global organizations operated under the quiet prestige of Project Glasswing, an early-access program that granted a handful of trusted partners a glimpse into the next frontier of machine intelligence. For the engineers at SK Telecom, this meant access to Claude Mythos, a model designed with a specialized, high-precision capability for identifying software vulnerabilities. It was a strategic windfall, positioning a South Korean telecom giant at the bleeding edge of AI-driven cybersecurity. But this week, the door slammed shut, not because of a technical failure or a contractual dispute, but because of a geopolitical calculation made in Washington.
The Sudden Blackout of Project Glasswing
The disruption began last Friday when the Trump administration issued a direct order to Anthropic, the AI startup behind the Claude series. The mandate was absolute: revoke access to Claude Mythos and Fable 5 for all foreign nationals, including immigrants residing within the United States. The scope of the order left Anthropic with a difficult operational choice. While the government demanded a filter based on nationality, Anthropic determined that implementing granular, nationality-based access controls would be nearly impossible from a privacy and technical standpoint. Rather than risk a partial or flawed implementation that could violate federal orders, Anthropic chose the nuclear option and completely disabled access to these models for all non-US users.
This move effectively dismantled the international wing of Project Glasswing. SK Telecom had been one of approximately 150 companies worldwide selected for the program after weeks of rigorous vetting and collaboration with US government officials and external experts. The models in question, particularly Claude Mythos, are not standard LLMs; they are highly controlled assets capable of identifying deep-seated software vulnerabilities that could be weaponized if they fell into the wrong hands. When the US government first flagged SK Telecom's access earlier this month, Anthropic complied immediately by revoking the company's credentials. However, that specific revocation was merely the precursor to a broader shift in policy that transformed a single company's access issue into a comprehensive export control measure for all foreign entities.
The Ghost of China Unicom and the Amazon Trigger
The reasoning behind the US government's aggression lies in a complex intersection of historical business ties and a recent technical failure. The primary catalyst was SK Telecom's legacy relationship with China. In 2004, SK Telecom established UNISK, a joint venture with the Chinese state-owned carrier China Unicom, to provide mobile content and wireless internet services. This partnership deepened in 2006 when SK Telecom invested 1 billion dollars into convertible bonds issued by China Unicom's Hong Kong subsidiary, securing a stake of approximately 6.6 percent.
On paper, these ties are largely dormant. SK Telecom exited the partnership in 2009, selling its stake for roughly 1.3 billion dollars. By 2024, the company's footprint in China had shrunk to a negligible scale, with annual revenues of about 1.9 million dollars and a staff of only seven employees. However, the US government views risk through a wider lens. The broader SK Group maintains extensive interests in China across the semiconductor and energy sectors, creating a perceived vulnerability. This perspective aligns with the recent trajectory of the Federal Communications Commission (FCC), which has proposed banning interconnections between US carriers and Chinese firms like China Unicom on the grounds of national security.
While the geopolitical tension provided the motive, a technical vulnerability provided the urgency. On June 9, researchers at Amazon discovered a critical flaw in Fable 5, the protected version of the Mythos model. The Amazon team found that they could bypass the model's safety guardrails, effectively unlocking the raw, potent cyber-attack capabilities of the underlying Mythos engine. This discovery was reported directly to the White House. The report served as a wake-up call for US officials, who concluded that Anthropic lacked the necessary safeguards to prevent the world's most powerful vulnerability-hunting AI from being misappropriated. The result was a swift transition from selective partnership to total isolation.
This incident reveals a harsh reality for the global AI ecosystem: capital investment does not equal guaranteed access. In 2023, SK Telecom invested 100 million dollars into Anthropic to build a commercial partnership focused on telecom-specific AI models. Yet, this massive financial commitment provided zero protection against a national security directive. The precedent is now clear that strategic alliances and equity stakes are secondary to the US government's export control lists and supply chain security standards.
For other Korean entities, the warning signs are flashing. Samsung Electronics and the Korea Internet & Security Agency (KISA) were also participants in Project Glasswing, meaning they are now subject to the same policy-driven disruptions. The reliance on a single, US-based frontier model has evolved from a technical choice into a systemic risk. When a single policy shift can instantly terminate research and disable services, the cost of dependency becomes unsustainable.
AI practitioners and enterprise architects must now pivot toward a strategy of model diversification. The era of betting on a single 'best-in-class' vendor is ending, replaced by a need for portable model portfolios that can be swapped instantly in response to regulatory volatility. Technical partnerships are no longer permanent fixtures; they are conditional permissions that can be revoked by a government agency at any moment.
