China's Mythos Moment

Anthropic's Claude Mythos rattled Washington enough that the White House built what amounts to a mandatory pre-approval system for frontier models, largely because of the model's knack for finding cyber vulnerabilities. The obvious question now is what happens when a Chinese lab hits that same capability ceiling — and how the Chinese Communist Party reacts when it does.

The timeline is getting shorter than people think. Back in June, Z.ai cofounder Jie Tang told Elon Musk on X that a Chinese model on par with Mythos would arrive before year's end, and the recent release of Kimi K3 suggests that estimate isn't far off. Reuters reported on July 7 that Chinese regulators are already weighing limits on overseas access to the country's most advanced AI systems. Xi Jinping used his WAIC speech to praise openness while also warning that Beijing must build out laws, technical monitoring, early-warning systems and emergency response plans to keep AI "under human control" and guard against misuse.

ChinaTalk's team decided to sketch out what that reckoning might actually look like, using a fictional Zhipu model called GLM 6 as a stand-in for whichever lab — Kimi, DeepSeek, Qwen, or someone else entirely — actually crosses the Mythos threshold first. They land on three plausible outcomes.

The first is the laissez-faire path: the model drops open-weight with little fanfare, cyber researchers quickly discover it can hunt zero-days the way Mythos can, enterprises pile in for the cheap performance, and Beijing has to weigh the economic upside of a dominant open model against the risk of proliferating cyberweapons before eventually tightening licensing after the fact. The second is a phased rollout modeled on Anthropic's Project Glasswing, where access starts with government ministries and trusted giants like Baidu, Alibaba and Tencent, expands slowly to provincial governments and other firms, and only reaches the public as a deliberately weakened version once bureaucratic caution and competitive pressure from abroad force the issue. The third is the black-box scenario, where the state claims exclusive access from day one, deploys the real model internally for security and defense purposes, and hands the public a lobotomized release while the full capability stays locked inside government and military channels — with the real firepower potentially used for cyber operations against less-prepared countries rather than acknowledged at home.

None of this is meant as prediction so much as preparation. The point, as the authors frame it, is to get policymakers, safety researchers and AI-policy people thinking through multiple branches of the tree now, rather than being caught flat-footed by whichever one actually happens.