A South Korean Telecom's Alleged China Ties Helped Trigger Washington's Crackdown on Anthropic's Mythos

When Anthropic pulled its two most powerful models offline late on the night of June 12, the company framed it as a regulatory accident. "We believe this is a misunderstanding and are working to restore access as soon as possible," it told customers, after Commerce Secretary Howard Lutnick sent a letter to chief executive Dario Amodei placing Claude Mythos 5 and Claude Fable 5 under export controls — barring their use not only abroad but by any foreign person on U.S. soil.

But reporting that emerged in the days that followed suggests the chain reaction did not begin with a model at all. It began with a single customer: SK Telecom, South Korea's largest wireless carrier, and the alleged China ties that set off alarms inside the White House.

How one partner became a national-security problem

According to Semafor, which first connected the dots on June 13, U.S. officials grew concerned after reviewing the roster of organizations Anthropic had granted early access to Claude Mythos — its frontier model for finding software vulnerabilities — through a partner program called Project Glasswing. Anthropic had submitted an initial list of roughly 111 organizations, which the administration reviewed and approved. A later, supplementary list, however, included a South Korean telecommunications firm that officials suspected of links to China.

The only Korean carrier publicly known to have joined Glasswing is SK Telecom, which announced earlier in June that it had secured early access to Mythos after joining the security consortium. The company invested $100 million in Anthropic in 2023 and built a commercial partnership to develop telecom-focused AI models.

The White House, according to the reporting from Semafor and German outlet The Decoder, told Anthropic to cut SK Telecom off. Anthropic complied immediately.

The China question

The concern traces to SK Telecom's corporate lineage rather than its present-day footprint. The carrier is part of SK Group, a sprawling conglomerate with substantial business in China, and SK Telecom itself held a stake in the Chinese state-owned carrier China Unicom until 2009. The company had bought roughly $1 billion in convertible bonds in 2006, converting them into a holding that was diluted to 3.8 percent before it sold out near the end of the decade.

SK Telecom's current China exposure is, by its own accounting, minimal: about $1.9 million in China revenue in 2024 and seven employees in the country, according to its annual report. The company denied any China connections in remarks to Korean media, and Korean outlets including the Korea JoongAng Daily and Kyunghyang Shinmun reported that the episode has rattled the broader Korean telecom sector, which now fears guilt by association in any future allocation of frontier-AI access.

These remain allegations. No public evidence has been presented showing that SK Telecom routed Mythos access to Chinese entities, and the company's denial is on the record.

The second shoe: Fable 5's flaws

The SK Telecom dispute might have stayed a quiet partner-vetting matter. What turned it into an export-control event was a second, separate incident. Within days, Amazon — and reportedly Amazon CEO Andy Jassy directly — flagged that security flaws in Claude Fable 5 could be used to bypass the model's safety restrictions, effectively jailbreaking guardrails on a system explicitly built to hunt vulnerabilities.

Stacked together, the two episodes — a partner Washington didn't trust and a model the administration feared could be turned against its own safeguards — appear to have cost Anthropic the benefit of the doubt. Lutnick's letter, later published by Bloomberg, warned of civil and criminal penalties for noncompliance and extended the restriction to all foreign persons inside the United States. It is the first time a U.S. frontier lab has been ordered to halt global access to specific models on national-security grounds.

Analysis: export controls as an access turnstile

For years, U.S. AI export policy targeted hardware — the chips that train and run large models. The Mythos episode marks a decisive shift to controlling the models themselves, and more pointedly, who is allowed to touch them. The instrument here was not a published rule but a letter, applied retroactively to a deployed product, triggered by the identity of one customer on a list.

That sets an uncomfortable precedent. If access to the most capable models now hinges on partner-by-partner vetting against opaque national-security criteria, frontier labs effectively become extensions of the export-control bureaucracy, and their commercial decisions become reviewable by the executive branch in near-real time. The criteria appear elastic: SK Telecom's China ties are largely historical and financially trivial, yet were sufficient to force a revocation. A standard that broad could ensnare any multinational with legacy exposure to China.

The allied-relations cost is real, too. South Korea is a treaty ally and a critical node in the U.S. semiconductor and AI supply chain. Singling out its largest carrier — and an Anthropic investor — on suspicion alone risks chilling exactly the kind of allied adoption Washington says it wants. Korean firms reading the news this week are not learning that they must avoid China; many already have. They are learning that the rules can change after the deal is signed.

What to watch

Anthropic has signaled it expects to restore Mythos and Fable access "in coming days," and one executive told Korean media the company is confident it can re-enable the models once the misunderstanding is resolved. The questions that matter now are structural. Will Commerce codify a formal vetting process for frontier-model access, or continue governing by letter? Will the SK Telecom revocation become the template for partner-level screening across OpenAI, Google, and others? And will allied governments — Seoul first among them — push back on a regime that lets Washington decide, customer by customer, who gets to use the world's most powerful AI.