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

Zhipu AI's introduction of China's first counterpart to Project Glasswing marks a shift in how the nation's leading artificial intelligence labs approach cybersecurity, according to an expert. The company, now known as Z.ai, unveiled its GLM-5.3 model and "Shield of Open Source" initiative on Friday. The program offers free security audits, automated code-auditing tools, and free model usage quotas for the open-source community.

Zhipu defended its decision, stating: "When the strongest spear is locked in the hands of a few, the best shield must belong to everyone." This initiative comes as there is a growing debate about whether open-weight models exacerbate cyber risks by lowering technical barriers for attackers. Zhipu pledged stricter oversight through a layered risk-review system for GLM-5.3, blocking high-risk requests while maintaining routine, low-risk developer tasks.

The company's "Cybersecurity Trusted Access" plan would reserve GLM-5.3's sensitive offensive capabilities for verified users. Gabriel Wagner, an AI governance researcher, sees Z.ai's approach as "Project Glasswing with Chinese characteristics," emphasizing openness as an asset rather than a drawback. This move contrasts with US labs like Anthropic, which have taken a more restrictive approach to access.

GLM-5.3 recently outperformed Anthropic's Mythos model on the CyberGym benchmark, making it the best general cybersecurity model, according to a researcher. However, Zhipu acknowledged that GLM-5.3 still lags behind US competitors in offensive capabilities. To address these gaps, Zhipu committed substantial resources to safety in its Chinese launch announcement, becoming one of the first Chinese developers to outline a distinct AI safety vision and citing safety concerns for delaying the release of GLM-5.3's open weights.