Korea's AI Coding Boom Collides With US Export Controls as NSA Testimony Complicates Restoration

South Korea's largest technology companies made a sweeping commitment to Claude Code this month, rolling it out across tens of thousands of engineers at NAVER, Samsung SDS, LG CNS, Nexon, and Hanwha Solutions. It was the most concentrated enterprise adoption of Anthropic's AI coding tool anywhere in Asia. The timing seemed fortuitous: Anthropic had just opened a Seoul office and announced that its Asia-Pacific revenue had grown more than tenfold in the past year, with Claude Code's weekly active users in Korea expanding sixfold over four months. But on day nine of the export control directive that followed, the full picture of why those top-tier models remained offline became clearer—and the reason had little to do with the narrow technical jailbreak that had triggered the ban.

On June 12, the Commerce Department sent Anthropic a directive under the Export Controls Reform Act of 2018, citing national security. The agency ordered the company to disable access to Fable 5 and Mythos 5—Anthropic's most capable models—for every user worldwide. The Korean deployments were unaffected because they ran on Claude Sonnet 4.6 and Opus 4.8, the older generation of models that the directive did not touch. But the ban itself was a blunt instrument. The legal doctrine invoked, known as "deemed export" under 15 CFR 734.13, treats releasing controlled technology to a foreign national inside the United States as equivalent to exporting it to that person's home country. The rule was written for semiconductor blueprints and weapons schematics—physical artifacts that can be inventoried and licensed individually. A cloud API endpoint serving hundreds of millions of users has no nationality field in its session tokens or request headers. There is no mechanism for Anthropic to verify, in real time, the citizenship of every user sending a query. The only way to comply was to shut off access for everyone.

Two separate developments had converged to produce the directive. First, the White House learned that SK Telecom, South Korea's largest wireless carrier and an Anthropic investor, had gained access to Mythos 5 through Project Glasswing—Anthropic's invite-only cybersecurity consortium for vetted organizations—and that US officials suspected undisclosed ties to China. SK Telecom has denied the allegations, and the evidence has not been made public. The White House asked Anthropic to revoke its Glasswing access, which Anthropic did immediately. Second, Amazon—Anthropic's largest investor, with a cumulative stake of approximately thirteen billion dollars—flagged a guardrail bypass in Fable 5 to the White House. Amazon researchers discovered that prompting the model to read a codebase and "fix this code" against known vulnerabilities surfaced those vulnerabilities, effectively bypassing the model's intent-classification layer. Security expert Katie Moussouris, founder of Luta Security and the only outside expert who reviewed the underlying research report, concluded that the capability exposed cannot be meaningfully fixed without weakening the model for legitimate defensive use—and that the same capability exists in other publicly available frontier models including OpenAI's GPT-5.5.

But the narrow jailbreak was not the full story. On June 21, ten days after the ban took effect, testimony surfaced that reframed the dispute as something larger. Senator Mark Warner, vice chair of the Senate Intelligence Committee, said NSA Director General Joshua Rudd told him that Mythos "broke into almost all of our classified systems, not in weeks, but in hours" during an authorized red-team exercise. The Economist's editor later clarified that the statement should not be read literally and depended on Mythos working alongside other tools in specific conditions. Warner used the example not to condemn Anthropic but to argue for mandatory pre-release government testing of frontier models, and to praise Anthropic for having submitted Mythos for controlled evaluation. Even with those caveats, the disclosure mattered. It confirmed that the US government's concern extended beyond the narrow "fix this code" jailbreak to the underlying autonomous offensive capability of Mythos-class AI. A model that can find and chain vulnerabilities at that speed, in authorized testing against its own government's infrastructure, is not a straightforwardly commercial product—regardless of the safety classifiers Fable 5 runs on top of it.

For the thousands of engineers at NAVER, Samsung SDS, LG CNS, and Nexon who started using Claude Code in the past two weeks, the practical situation was this: the tools they were using worked. Claude Code, Claude Cowork, and the full Opus 4.8 and Sonnet 4.6 lineup were fully available. What they did not have access to was the top-tier capability Anthropic had positioned as the leading edge of autonomous coding and cybersecurity work. During Fable 5's brief availability, Stripe reported that the model completed a codebase-wide migration of a fifty-million-line Ruby codebase in a single day—work that would have taken a full engineering team more than two months by hand. That benchmark captured the gap between what Korean engineers were working with now and what they had been told was coming. Anthropic's international managing director Chris Ciauri told reporters at the Seoul opening that he was "very confident" the models would return within days. Prediction markets gave roughly a fifty-seven percent probability of restoration before July 1 as of June 18. As of June 21, no restoration had been announced, and the NSA testimony circulating that weekend made the near-term timeline harder to call.

The deepest structural problem was not a question of Anthropic's intentions or SK Telecom's affiliations. It was a question of whether the legal architecture that governs the export of dual-use technology was built to handle a capability delivered at consumer scale through a shared cloud endpoint. Export control law developed around physical goods that can be tracked, licensed, and recovered. It then extended, imperfectly, to static software and technical data—artifacts that can be inventoried and restricted by recipient. Live AI inference over a cloud API has none of those properties. A query to Fable 5 and a query to Claude Sonnet 4.6 look identical to Anthropic's infrastructure until the model responds. There is no chokepoint at which nationality can be verified, no artifact that can be recalled, and no mechanism for partial compliance short of a global shutdown. That is what happened on June 12: a legal framework designed for another era of technology produced a global outage to enforce a domestic restriction—and the first time in history a US government directive pulled a commercially deployed frontier AI model for every user on earth. Whether the right answer is a better enforcement mechanism or a categorical access model for autonomous offensive AI capability, both answers take time. Korean enterprises, Anthropic, and their mutual commercial relationship are now waiting on that question.