Zhipu AI’s answer to Project Glasswing marks shift for Chinese cyber safety: researcher

Zhipu AI’s introduction of China’s first answer to Project Glasswing marks a shift in how the country’s leading AI labs tackle global cybersecurity, according to a researcher. The Beijing-based company, now known globally as Z.ai, unveiled its GLM-5.3 model alongside the launch of its "Shield of Open Source" initiative. The programme offers free security audits, automated code-auditing tools and free model usage quotas to the open-source community.

Zhipu emphasized that when powerful tools are controlled by a few, the best defense should be accessible to everyone.

The initiative arrives amid a growing global discussion on whether open-weight models exacerbate cyber risks by lowering barriers for attackers. Zhipu pledged stricter oversight with a layered risk-review system for GLM-5.3, which blocks high-risk requests without hindering routine, low-risk developer tasks. Sensitive offensive capabilities of the model would be restricted to verified users under a "Cybersecurity Trusted Access" plan.

Gabriel Wagner, an international AI governance researcher, noted that Z.ai is proposing a "Project Glasswing with Chinese characteristics" where openness is an asset rather than a drawback. This approach differs from US labs that focus on restricting access to a limited number of players. Zhipu claimed GLM-5.3 outperformed Anthropic’s Mythos 5 in identifying and validating vulnerabilities on the CyberGym benchmark, making it the best model for general cybersecurity tasks.

However, Zhipu acknowledged that GLM-5.3 still lags behind US competitors in offensive capabilities, scoring lower than Mythos and OpenAI’s GPT-5.6 Sol on ExploitBench. Despite these shortcomings, Zhipu dedicated a significant portion of its launch announcement to safety and risk mitigation. The delay in releasing GLM-5.3’s open weights for up to two weeks, while completing security-hardening procedures, demonstrated the company's commitment to restricting its offensive capabilities while preserving its defensive value.

This rare instance of a Chinese developer taking the initiative to address AI safety signals a more sophisticated approach to open-weight risk management in China.