US forced Anthropic to cut SK Telecom over China ties

Key Takeaways

- The White House ordered Anthropic to cut SK Telecom's access to its Claude Mythos model over alleged China ties
- SK Group held a stake in state-owned China Unicom until 2009 and maintains major Chinese business interests
- Security flaws in Fable 5 discovered by Amazon compounded the crisis, leading both models to be taken offline
Anthropic shut down two of its AI models, Claude Mythos and Fable 5, after the White House intervened over alleged Chinese government connections at a South Korean partner. The trigger: SK Telecom's access to Mythos through Anthropic's partner program, Project Glasswing.
According to WIRED, US officials grew alarmed at what they viewed as SK Telecom's ties to China. The White House told Anthropic to revoke SK Telecom's access. Anthropic complied immediately.
SK Telecom denied any China connections to a Korean newspaper. But the denial sidesteps a structural reality: SK Telecom belongs to SK Group, South Korea's third-largest conglomerate with roughly $220 billion in annual revenue. The group maintains extensive business operations in China spanning semiconductors, energy, and telecommunications.
What is SK Group's history with China?
SK Group held a stake in China Unicom, one of China's three state-owned telecom carriers, until 2009. That ownership history matters. Even divested, the relationship points to deep commercial ties with entities under Beijing's control.
For US national security officials scrutinizing who touches frontier AI systems, past connections can be as concerning as current ones. The question isn't just who owns what today. It's what access paths remain open.
Why did the crisis escalate to a full shutdown?
Cutting SK Telecom's access might have contained the problem. It didn't. Shortly after, Amazon and other companies flagged security vulnerabilities in Fable 5. These flaws could reportedly bypass the model's safety restrictions.
Two incidents in quick succession proved fatal to White House confidence in Anthropic. First, a partner access issue tied to China. Second, fundamental security gaps in a production model. The administration forced both Claude Mythos and Fable 5 completely offline.
The speed of the shutdown underscores how little margin for error exists in frontier AI. Anthropic has built its brand on safety. Dario Amodei, the company's CEO, has called AI development "one of the most transformative and potentially dangerous technologies in human history." When the company that warns loudest about AI risks trips over partner vetting and security flaws, regulators notice.
What does this mean for AI partner programs?
Project Glasswing suggests Anthropic was expanding access to its most advanced models through partnerships. That model now carries obvious risks. Any partner with material business interests in China, or historical ties to state-owned enterprises, could trigger similar interventions.
South Korean tech companies are caught in an uncomfortable position. SK Telecom has over 30 million mobile subscribers and has publicly stated ambitions to become "a global AI company." But South Korea's geographic and economic realities mean deep commercial ties to both the US and China. Those ties now conflict.
For AI companies, the lesson is blunt: national security vetting of partners must extend to corporate structures, historical relationships, and even divested stakes. The US government is treating frontier AI access as a security matter, not a commercial one.
What happens to Anthropic now?
Anthropic retains significant backing from Google and Salesforce, with a valuation around $2 billion as of early 2024. The company's reputation rests on responsible AI development. Losing two models to a White House intervention damages that standing, regardless of whether Anthropic ultimately handled the crisis correctly.
The company cooperated immediately with the access cutoff. That compliance may matter for future relationships with regulators. But it also means Anthropic had to sacrifice models and partnerships at government direction. What looked like a partner program became a national security liability overnight.
DeepMind's framework for treating AI systems as potential insider threats connects directly to the security concerns that drove this Anthropic shutdown.
Frequently Asked Questions
Why did the US government force Anthropic to cut SK Telecom's access?
US officials were concerned about SK Telecom's alleged ties to China, including its parent company SK Group's extensive Chinese business interests and historical stake in state-owned China Unicom.
What is Project Glasswing?
Project Glasswing is Anthropic's partner program through which SK Telecom gained access to the Claude Mythos model before being cut off at White House direction.
What security flaws were found in Fable 5?
Amazon and other companies identified vulnerabilities in Fable 5 that could be exploited to bypass the model's safety restrictions. The specific technical details have not been disclosed.
Are Claude Mythos and Fable 5 permanently offline?
The models were forced completely offline by the White House after the dual incidents. Whether Anthropic can bring them back depends on addressing both the partner access and security concerns.
Did SK Telecom actually have China ties?
SK Telecom denied China connections, but its parent company SK Group held a stake in China Unicom until 2009 and maintains significant business operations in China.
Logicity's Take
This incident marks a turning point for AI industry partnerships. Companies building frontier models now face implicit government vetting of their entire partner ecosystem. The bar isn't just whether a partner is trustworthy today, but whether any corporate ancestor or affiliate has touched Chinese state enterprises. For South Korean, Japanese, and European tech companies with legitimate business in China, this creates a binary choice that didn't exist two years ago.
Need Help Implementing This?
AI partnership vetting and compliance frameworks are becoming essential for technology companies. Contact Logicity's advisory team to understand how geopolitical risks affect your AI strategy and partner relationships.
Source: The Decoder / Matthias Bastian
متطلبات جديدة لمشاركة البيانات: Anthropic تفرض الاحتفاظ ببيانات المستخدمين لمدة 30 يوماً
المقال الجديد يكشف عن متطلبات تقنية جديدة: استخدام Claude Fable 5 أو Mythos 5 على Amazon Bedrock يتطلب الآن مشاركة بيانات الاستدلال مع Anthropic لمدة 30 يوماً مع مراجعة بشرية، وهو تغيير جوهري عن النماذج السابقة التي كانت البيانات فيها تبقى داخل حدود AWS. Anthropic تبرر هذا كمتطلب أمني لنماذج فئة Mythos للكشف عن الهجمات الجديدة.
Huma Shazia
Senior AI & Tech Writer
Related Articles
Browse allZuckerberg's Superintelligence Lab Faces Setback
The first AI model from Zuckerberg's superintelligence lab has failed to impress compared to its rivals, sparking concerns about the lab's direction. We take a closer look at what happened and why it matters.

Muse Spark Launch Propels Meta AI App to Top 5
The recent launch of Muse Spark has significantly boosted the popularity of Meta AI app, pushing it into the top 5. We explore what this means for the AI landscape.

Meta's Muse Spark AI Model Lags Behind ChatGPT and Claude
Meta's Muse Spark AI model still can't outperform ChatGPT and Claude in key areas, despite its advancements. We explore what this means for the AI landscape.

Meta Launches Muse Spark AI To Challenge ChatGPT
Meta launches Muse Spark AI to challenge ChatGPT and Claude, we explore what this means for the AI landscape. Muse Spark AI is a significant development in the AI chatbot space.


