The rapidly evolving landscape of artificial intelligence has hit a major geopolitical snag, centered on the San Francisco-based AI lab Anthropic. At the heart of the controversy is “Claude Mythos,” a cutting-edge AI model so advanced that it possesses the unsettling ability to identify and exploit complex software vulnerabilities. Because of this, Anthropic’s distribution strategy was exceptionally cautious, utilizing a program called “Project Glasswing” to gatekeep access to only about 150 trusted organizations. However, the situation spiraled when the Trump administration raised alarms over one specific participant: South Korean telecommunications giant SK Telecom. US officials expressed deep-seated anxiety regarding the company’s alleged ties to China, fearing that the specialized cyber-capabilities of Mythos could inadvertently slip into the wrong hands. Despite SK Telecom’s firm denial of any such affiliations, the White House remained unconvinced, setting the stage for a high-stakes standoff between executive power and private sector innovation.
The tension escalated significantly when a second, seemingly unrelated problem emerged from the tech giant Amazon. After Anthropic released “Fable 5″—a public-facing, “guardrailed” version of the Mythos model on June 9—Amazon’s internal security researchers discovered cracks in the system. They reported that they could effectively bypass the safety protocols, potentially exposing the raw, formidable power of Mythos to users who shouldn’t have access to it. While Anthropic and various cybersecurity experts have argued that these vulnerabilities are essentially “par for the course” in the current state of AI development—not unique to Claude—the timing could not have been worse. To the Trump administration, the combination of SK Telecom’s international footprint and Amazon’s vulnerability report signaled that Anthropic’s safety framework was insufficient, leading to a profound loss of trust in the lab’s ability to police its own technology.
Under mounting pressure, the White House took a decisive and controversial step: it ordered Anthropic to revoke access to its most powerful models for all foreign nationals, a mandate that would include immigrants currently residing and working within the United States. Anthropic, facing the logistical impossibility of building a privacy-preserving system that could effectively vet the nationality of every single user without shattering their commitment to user data protection, made a radical choice. Rather than implementing these invasive gatekeeping measures, the company decided to essentially flip the kill switch. They disabled access to the Mythos and Fable 5 models entirely, effectively putting their most sophisticated research on ice while they wrestle with federal regulators in private.
The fallout has left the industry in a state of nervous speculation. Sources close to Anthropic have attempted to separate the two issues, noting that the official government letter demanding access restrictions made no specific mention of the South Korean telecommunications company or China. They suggest that the government’s demands were broad, sweeping, and perhaps disconnected from the specific security flaws Amazon had highlighted. Nevertheless, the reality for the AI community is that the world’s most potent “white hat” hacking AI is currently offline. The breakdown in communication between the White House and lead researchers at Anthropic remains a glaring example of the friction between the frantic pace of AI development and the slow, heavy hand of national security policy.
The connection to SK Telecom is particularly complex, illustrating how globalized the modern tech economy truly is. While the mobile carrier itself operates largely within South Korea, it is a subsidiary of the massive SK Group conglomerate, which maintains sprawling tentacles across the Chinese semiconductor and energy sectors. Washington’s apprehension is rooted in a “guilt by association” framework: even if a company like SK Telecom is a trusted partner to Anthropic, the potential for technology to flow through the wider corporate architecture into China is a risk the government currently refuses to ignore. Projects like “Project Glasswing”—intended to foster collaborative security-building with entities like Samsung and Korea’s security agencies—now appear to be casualties of an era where global cooperation is increasingly viewed through a lens of defensive nationalism.
Ultimately, this saga highlights the precarious middle ground that AI labs occupy today. They are tasked with developing technology that could revolutionize software security, yet that same potential makes them prime targets for export control and intense regulatory scrutiny. As negotiations between the White House and Anthropic drag on, the incident serves as a warning to the entire tech sector: the era of open-ended AI experimentation is being quickly replaced by a new, stricter paradigm of sovereign control. Whether these models will ever return to the public sphere without being kneecapped by draconian access limitations remains an open question, leaving researchers, commercial partners, and the global security community waiting to see who holds the final authority over the digital tools of the future.
