Korea falls back on homegrown AI after White House bans Mythos exports

NEWS ANALYSIS

In response to the White House's directive blocking exports of Anthropic's top-tier AI models, Korea is advancing two domestic AI initiatives: one to bolster AI security through the white-hat hacker community, and another to develop homegrown foundation models for critical public-sector and national security applications.

The ban came just 10 days after state-run Korea Internet &  Security Agency (KISA) and tech giants Samsung Electronics and SK hynix secured access to the Mythos AI model under Anthropic's Glasswing Project.  The Anthropic-led project drew attention for who would be brought into the exclusive circle of users of Claude Mythos Preview — Anthropic's unreleased frontier model — for defensive cybersecurity purposes, finding and fixing software blind spots in key infrastructure before malicious actors can exploit them. 

Now that this has been upended, the two local projects have become the main response to the AI export ban. However, one is still in its early stages while the other faces a widening capability gap between homegrown and U.S. models. The stark irony is that most of the industry heavyweights participating in these initiatives are simultaneously expanding their use of third-party AI, such as GPT and Claude, in their own systems, and some consortia tasked with building a homegrown foundation model were caught using foreign open-source models instead. 

Nonetheless, Korean experts say the Mythos ban should be treated as a warning sign to accelerate AI infrastructure expansion and protect domestic data, which underpins AI training and should not be routed through foreign models. 

A Canopy to shelter Glasswing

Korea's cybersecurity community, which had been quietly preparing for the worst since Anthropic announced Project Glasswing, announced a response within days of the Mythos blockage.

The initiative, called Project Canopy, is a civic AI security effort launched on June 17 by Project Plasma, a nonprofit white-hat hacker organization, with 27 participating companies and institutions, including LG Electronics, Hyundai Motor Group, NHN and SK.

Project Canopy is designed to safeguard sensitive data held by public institutions, including hospitals, schools and government agencies, by providing AI-powered cybersecurity support. Those institutions are considered among the most vulnerable targets for data breaches and AI-powered cyberattacks, yet many lack the financial resources to defend themselves.

“When we used to warn that the United States could suddenly cut off API access, people treated it as a hypothetical,” said a source in the IT industry who requested anonymity. “Now that it’s actually happened, most countries are scrambling, but Korea moved early enough that we’re still able to create a response team.”

The project runs three programs. One distributes free AI-powered vulnerability scanning credits to open-source maintainers. Another offers AI-powered security scans to public institutions — government agencies, courts, local governments, hospitals and schools. The third offers rewards to white-hat hackers who verify AI-flagged vulnerabilities and follow through on deploying patches.

Seed funding comes from Theori Korea, a Korean-led cybersecurity startup based in Austin, Texas, contributing roughly 3 billion won ($1.92 million) in credits for its AI security analysis tool, Xint.

A separate government-backed initiative, led by the Korea Information Security Industry Association, also operates under the K-Glasswing name as a training program for cybersecurity professionals.

Still, such efforts may not be enough to match Mythos — not for lack of technical skill, but because no domestic frontier model yet exists that can scan millions of lines of code, spot weaknesses in real time, and match the speed and scale at which a model like Mythos operates. That is why experts have been calling for a native AI model that could become the backbone of the country's public and security operations. 

"When it comes to data and systems tied to national security, there are clear limitations on relying on foreign AI to do the work," said Kim Seung-joo, a professor at Korea University's School of Cybersecurity. "That is why defense, diplomacy and public systems must move toward an independent AI infrastructure developed domestically, one that cannot be suddenly cut off by a foreign counterpart. Feeding national defense or diplomatic data into foreign AI also means that data passes through overseas servers, no matter how strong the contractual protections are."

Data matters more than an AI arms race

Korea had already seen this coming. Since August 2025, the government has opened a nationwide competition gathering technology firms to build a foundation model for public and defense purposes, envisioning a sovereign AI of its kind, which refers to a country's capacity to develop, own and control its own AI infrastructure without relying on foreign providers.

The government's target is for the winning model to perform at 95 percent of the level of top global AI models released within the previous six months. OpenAI and Google were the original benchmarks, with Anthropic added as a reference point against which competing Korean firms are now measured. 

The competition launched with ambitions that quickly ran into friction when three out of five groups were alleged to have copied or borrowed frameworks from Chinese models. Naver Cloud was disqualified after its vision encoder was found to share similarities with Alibaba's Qwen. AI startup Upstage was also forced to defend the originality of its model through a livestream audit after rumors spread that its Solar model had been cloned from Zhipu AI's GLM. SK Telecom faced similar accusations over alleged similarities to DeepSeek.

The disqualification of Naver Cloud, which was expected to be a finalist in this competition, prompted the government to add another competitor, Motif Technologies.

Final winners are expected to be selected in early 2027. Skeptics doubt whether the government-led project can produce a model on par with U.S. or Chinese counterparts, but local AI experts push back on that framing.

"Korea does not need a foundation model that can go toe-to-toe with foreign counterparts to make sovereign AI worthwhile, because the bar for public services and national security applications is different from the threshold for general consumer AI," said Choi Byung-ho, a research professor at Korea University's Human-Inspired AI Research Center. A less capable model than GPT or Gemini, he added, can still run most government systems.

"Where raw capability matters more, the gap can be bridged through architecture," he continued. "The emerging approach is to build multi-agent systems around the base model, layering specialized AI agents on top so that complex, multi-step tasks get routed to the right tool at the right time. Korea's foundation model doesn't have to do everything brilliantly on its own. It serves as the sovereign core, the piece that Korea owns and controls, while agents built around it extend its reach into more demanding applications."

BY LEE JAE-LIM [[email protected]]