Ten days ago, the US government gave Anthropic 90 minutes to pull its two most capable models offline, citing a national security risk. This week, the same company opened a new office in Seoul and announced that Samsung, LG, NAVER, Hanwha and Nexon were all deploying its AI across their workforces. Both things are true at once, and the gap between them is the most revealing thing about how AI is now being governed.
The freeze itself appears to be thawing. At the G7 summit on Wednesday, President Trump said negotiations with Anthropic were "going fine." By Friday he told Axios he no longer regarded the company as a national security threat, adding only: "Well, not now, but a week ago, maybe." That is a remarkable walk-back for a designation that, a week earlier, had been serious enough to bar some of Anthropic's own employees from using its software. Chris Ciauri, the company's managing director of international, struck the same note in Seoul: "We are very confident that in the coming days, the models will become available again."
The thaw matters less for its timing than for what it exposes. An export control that can be imposed in 90 minutes and softened by an offhand interview answer is not a framework; it is a mood. The government has told Anthropic it must eliminate every jailbreak from Fable 5 before relaunch, a bar that security researchers across the field describe as technically impossible. New attack techniques arrive faster than any lab can enumerate them, a point Anthropic made in its own launch documentation. The likely real resolution is more modest: proactive testing and government notification, built on monitoring tools Anthropic already runs.
While Washington decided whether to trust the company with its own technology, Korean enterprises were busy betting their engineering operations on it. NAVER, the country's largest web portal, has rolled out Claude Code across its entire engineering organisation, even though it builds a rival model of its own. Samsung SDS and LG CNS are pushing Claude into hundreds of thousands of employees. Anthropic signed a memorandum with Korea's science ministry on AI safety the same week. For a company days from an October IPO filing, it was the most commercially significant day in its Asia-Pacific history.
The contradiction is not really a contradiction. A model capable enough to alarm a national security apparatus is, for the same reasons, capable enough to be worth deploying across a conglomerate. The open question is whether enterprise customers, watching a flagship model vanish on a government letter and return on a presidential shrug, will keep treating any single closed model as something they can build on. The next story on this page suggests some of them have already drawn their own conclusion.