White House Revoked SK Telecom's Anthropic Mythos Access

Wired identified SK Telecom as the South Korean carrier whose access to Anthropic's Claude Mythos was revoked at White House request in early June. The revocation arrived days before the Commerce Department's June 12 directive disabled Mythos 5 and Fable 5 for all foreign nationals. The two actions compounded: every Korean enterprise dependent on Anthropic's frontier models lost access overnight.

SK Telecom joined Project Glasswing — Anthropic's invite-only cybersecurity consortium for controlled Mythos access — in a second-wave expansion in early June that brought total membership to roughly 150 organizations. Anthropic submitted an initial list of 111 Glasswing partners to the Trump administration for review; that list was approved. SK Telecom appeared on a supplemental submission. The White House flagged it and asked Anthropic to revoke the carrier's credentials. Anthropic complied the same day with no export-control threat issued, per Wired's reporting.

The national-security concern centered not on SK Telecom's direct China operations—$1.9 million in Chinese revenue in 2024, seven employees there—but on its parent SK Group's extensive interests in Chinese semiconductors and energy. SK Telecom formed a wireless joint venture called UNISK with state-owned China Unicom in 2004, invested $1 billion in China Unicom convertible bonds in 2006 (which converted to approximately 6.6%), and sold that stake for $1.3 billion in 2009. A residual UNISK investment worth roughly $17 million appears in SK Telecom's 2025 SEC filing. SK Telecom, which invested $100 million in Anthropic in 2023, denied the allegations. A Korean newspaper reported the company said the anonymous claims "lack verified facts."

The June 12 ban traces to a separate trigger. Amazon, Anthropic's largest investor with approximately $13 billion cumulative stake, flagged a guardrail bypass in Fable 5, Anthropic's public-release version of the Mythos-class capability. Researchers found that prompting the model to "fix this code" turned it into a vulnerability-discovery tool despite an explicit guardrail against that use. Amazon CEO Andy Jassy raised the findings directly with White House officials. The Commerce Department order followed within hours. Rather than implement nationality-based access filtering, Anthropic disabled both Mythos 5 and Fable 5 for all customers globally.

Korean enterprises running standard Claude remained unaffected: Naver (Claude Code across its engineering organization), LG CNS (deployed to thousands of employees), and Samsung SDS (Claude Cowork and Claude Code for Samsung Electronics staff) use generally available Claude models, not Glasswing-gated Mythos. The disruption hit specifically where Anthropic granted access to its most capable tier. Samsung Electronics, SK Hynix, and the Korea Internet & Security Agency (KISA) were also Glasswing participants; all Korean institutional access was revoked by the June 12 directive. Anthropic says restoration will come through "Glasswing Wave 2."

Glasswing is a closed API program. Revoking credentials requires no public announcement, regulatory process, or notice to end users. The entire enforcement action—government request to private company, credential revocation, compliance confirmation—ran inside the commercial relationship. Export enforcement happened between companies, not at a border. This creates operational risk for any team running production pipelines on restricted-tier model access: the revocation surface is a private dashboard, not a published sanctions list.

Approximately 100 cybersecurity professionals, including former Facebook security chief Alex Stamos and Luta Security's Katie Moussouris, signed an open letter arguing that Mythos-class models, while "quite good" at finding and weaponizing software flaws, are "not uniquely good." They called for the controls to be lifted. Mythos 5 and Fable 5 remain offline. Anthropic's international chief said restoration was "days" away as of June 18.

If your production stack runs on a vendor's controlled-access tier rather than its generally available API, model continuity is now a geopolitical dependency. Treat it like a single-region deployment and design the fallback before you need it.