US Blocks Foreign Access to Anthropic's Top AI Models, Citing Security Risks; Korea's 'Trusted Partner' Status in Flux

The U.S. government has taken the unprecedented step of ordering AI company Anthropic to block all foreign nationals from accessing its most powerful models. This move marks a significant turning point in AI governance, as it's the first time that providing access to an AI model has been treated as a form of 'export' subject to national security controls. The immediate target of this order was Anthropic's 'Project Glasswing,' a program designed to give trusted partners early access to its frontier models, like Mythos 5, to find and fix security flaws. For Korean companies and institutions that were part of this program, access was cut off overnight. What caused this sudden policy enforcement? The core issue appears to be a breakdown of trust between the White House and Anthropic over its partner vetting process. According to a Washington Post report, the U.S. government was alarmed to find that a South Korean telecommunications company with alleged ties to China was on the Glasswing access list. This incident wasn't isolated; it followed months of growing friction, including a formal designation by the Pentagon labeling Anthropic a 'supply chain risk' earlier in the year. This event is also a major test for the Trump administration's evolving AI export policy. After rescinding a broader, more systematic rule that controlled the export of AI model weights, the administration shifted to a more targeted, case-by-case approach. This strategy aims to promote the export of American AI technology to allies under a 'Trusted Partners' framework while selectively blocking access for others. The Anthropic case, triggered by a specific security lapse, has become the first real-world application of this ad-hoc control mechanism. For South Korea, the implications are significant. The market's reaction was telling: while semiconductor giants like SK Hynix continued to rally on the global AI hardware boom, SK Telecom's stock fell sharply, reflecting the direct risk to its business from losing AI model access. This highlights that even for a key U.S. ally, being a 'Trusted Partner' is not guaranteed. Access to critical technology now depends on rigorous scrutiny of supply chains and corporate governance. The path forward now hinges on negotiations to define what it means to be a 'Trusted Partner.' While recent G7 discussions suggest a willingness to create a licensed access path for allies, the criteria will undoubtedly be far stricter, demanding robust security protocols and verifiable independence from strategic rivals. - Deemed Export: A U.S. legal concept where transferring controlled technology or information to a foreign national within the U.S. is considered an 'export' to that person's home country. - Project Glasswing: An Anthropic initiative providing early, controlled access to its most advanced AI models for security researchers and trusted partners to identify and mitigate risks. - EO 14320: An Executive Order establishing the 'American AI Exports Program,' which aims to promote the export of U.S. AI technology to allies and partners while implementing security controls.