As policymakers grapple with the governance of advanced AI models, a Chinese open-weight model has narrowed the gap with industry leaders like OpenAI’s GPT-5.6 Sol and Anthropic’s Mythos. GLM-5.2 from Z.ai is only a few months behind on cyber and bio capabilities, according to SaferAI's report.

However, the safety divide widens: while GLM-5.2 refused no offensive tasks, Claude Opus 4.7 consistently rejected them, rendering some evaluations impossible. This highlights the risk of powerful AI falling into the wrong hands with no means of policing use post-download.

The debate now shifts from competition to managing risks. Open-weight models are rapidly closing in on frontier capabilities, making it imperative to address safety measures. While closed models rely on safeguards like classifiers and refusal training, these become ineffective once weights are run locally where any protection can be bypassed or removed.

Frontier developers have started employing selective restrictions and pre-deployment safety evaluations, but the effectiveness of these remains questionable. As China acknowledges AI risks, its regulations focus less on catastrophic threats than on political content and social stability.

This shift underscores the need for a global consensus on AI governance to ensure that these powerful tools are used ethically and responsibly.