How export controls and geopolitics blocked Anthropic's Mythos rollout with SK Telecom as a focal point

Anthropic took its most advanced AI models offline after the White House ordered access revoked for all foreign nationals, a move triggered by concerns over SK Telecom's alleged ties to China and Amazon-flagged vulnerabilities in the safeguarded version Fable 5. For AI builders and product teams, this episode is a practical warning: geopolitics and export controls can strip access to frontier models overnight, even for trusted international partners.

What happened

U.S. officials alleged that SK Telecom, South Korea's largest wireless carrier and a $100 million investor in Anthropic, had ties to China. Those concerns surfaced after Anthropic granted SK Telecom access to its cutting-edge Claude Mythos model through a controlled early-access program called Project Glasswing. Around 150 companies received Mythos access, including Samsung Electronics and the Korea Internet & Security Agency.

Separately, Amazon researchers reported that Fable 5, a highly safeguarded public version of Mythos, had vulnerabilities that could be circumvented to reach Mythos's cyber capabilities. Anthropic and external security experts maintain these risks are not unique to Claude, but the combination led the White House to conclude it could not trust Anthropic to safeguard its most powerful AI. The administration ordered Anthropic to revoke Mythos and Fable 5 access for all foreign nationals, including immigrants inside the U.S. Rather than gate by nationality, Anthropic chose to disable access entirely. Negotiations with the White House remain ongoing without service being restored.

Why AI builders should care

If your team relies on early access to frontier models through trusted partnerships, this is a concrete example of how quickly that access can disappear. Policy actions targeting a single partner or a perceived vulnerability can cause the model owner to pull the plug for everyone international. For builders running agent workflows, security research, or product features atop these models, any interruption means lost development time, broken integration testing, and potential product delays.

Beyond immediate access, the controversy shows that investments and partnerships across borders become geopolitical flashpoints. SK Telecom's $100 million stake and participation in Project Glasswing were seen as assets until they triggered scrutiny. Any foreign entity with ties to China, even through a parent conglomerate like SK Group, can become a target for export controls.

Practical implications

Builders should map the policy exposure of their international AI partnerships before a crisis hits. This includes understanding the nationality and ownership structures of model access partners, and having fallback plans if access is suddenly restricted. For product teams deploying models in multi-region environments, export-control readiness should be part of governance checklists.

Guardrails and security assessments are no longer just technical issues; they are central to policy debates. If a model has proven vulnerabilities, even a safeguarded version can trigger government action. Teams building on top of guarded models should monitor public vulnerability disclosures closely. Finally, anticipate that future early-access programs may include more restrictive nationality screening, which could affect how you build multilingual or global features.

Caveats

This reporting relies heavily on a single in-depth WIRED investigation, supplemented by industry coverage. The exact timeline and internal government decisions may evolve as policy discussions continue. SK Telecom has publicly denied having ties to China, and Anthropic views the SK Telecom access and the Fable 5 vulnerabilities as separate issues. The White House and Anthropic had not resumed service at the time of publication, and negotiations remain fluid. Claims from Amazon about Fable 5 vulnerabilities are contested by Anthropic and some security experts.