The Washington Post reported, and the Korea JoongAng Daily and Korea Times confirmed in subsequent coverage on June 18–19, 2026, that the trigger for the US export-control directive disabling Anthropic's Fable 5 and Mythos 5 for all foreign nationals was a South Korean telecommunications company with historical commercial ties to China. The carrier had been granted access to Mythos 5 through Anthropic's Project Glasswing, the company's vulnerability-discovery program that enrolled roughly 50 partners to use the model on critical infrastructure code. Anthropic International Managing Director Chris Ciauri said at the Seoul office launch that he was "very confident" the models would return "in the coming days." As of this digest, no restoration has been announced.
The reveal changes the shape of the story. The June 12 ban was originally framed as a generic concern about jailbreaks of vulnerability-discovery capabilities; the specific trigger now turns out to be a partner-vetting failure inside Anthropic's own program. Project Glasswing was designed to put a frontier model in front of critical software early enough to find vulnerabilities before adversaries did — admitting a partner with potential China exposure inverted the threat model. That is not a model-safety failure; it is an enterprise-security failure, and a different category of fix.
The downstream effects are already visible in adjacent stories from the same 24 hours. Swiss Prem AI confirmed a $100M Series A pitched explicitly at the on-premise alternative to Anthropic and OpenAI; Dream closed $260M at a $3B valuation selling sovereign-grade AI to governments; Anthropic's own Korean enterprise announcements last week now sit awkwardly next to the news that a Korean partner inside its safety program is what set off the directive. Allied governments are watching how Anthropic resolves this, because the precedent it sets — who can access frontier safety programs, under what controls — will define the rules for every comparable program at OpenAI, Google DeepMind, and xAI.
Takeaway for learners: partner programs are part of the security perimeter. If you ship a model — or any high-leverage tool — to fifty organizations under a vulnerability-discovery banner, your weakest partner is your real attack surface. The lesson for builders is unglamorous: vetting, contracting, and ongoing monitoring of program participants are not the boring part of safety work, they are the part that decides whether the program survives a single bad day.