China-based AI firm Zhipu has launched its latest flagship model GLM-5.3, surpassing Anthropic’s Mythos 5 in a key cybersecurity benchmark, marking a pivotal advancement in China’s strategic push to strengthen AI-driven cyber defence capabilities.
- GLM-5.3 outperforms Mythos 5 in identifying code vulnerabilities
- Model to be open-sourced for global cybersecurity benefit
- Competitive AI race pushes China’s cyber defence edge
What happened
Chinese AI company Zhipu (Z.ai) introduced its new model GLM-5.3, achieving an 84.5% success rate on CyberGym, a benchmark that evaluates AI models’ capacity to find and verify security weaknesses in source code. This score slightly exceeds those of Anthropic’s Mythos 5 and OpenAI’s GPT-5.6 Sol models, marking a significant milestone for China's AI capabilities in cybersecurity.
Despite its higher score on vulnerability detection, GLM-5.3 scored lower on ExploitBench, a test measuring exploit generation, indicating room for improvement in offensive cybersecurity applications. The firm plans to open source the model’s weights following a safety review, aiming to democratize AI-based cyber defence tools globally.
Why it matters
GLM-5.3’s advancements highlight China’s intent to match or surpass Western AI powers in cybersecurity, enhancing national defensive capabilities amid growing digital threats. By releasing the model openly, Zhipu signals a commitment to global collaboration rather than exclusive national control over critical cyber defence technologies.
This move could shift the cybersecurity landscape by making advanced AI-driven vulnerability detection accessible worldwide, potentially accelerating innovation while raising complex questions about AI ethics, security, and dual-use risks. The competition between Chinese and Western AI labs reflects broader geopolitical dynamics impacting technology development and deployment.
What to watch next
The upcoming public release of GLM-5.3’s parameters will be closely observed by the global cybersecurity and AI research communities to assess how widely and effectively the tool can be adopted or improved upon. Its real-world impact will depend on ongoing validation and safety measures to prevent misuse.
Further innovation is expected as Chinese AI firms continue scaling model parameters and refining capabilities. The evolution of exploit-generation benchmarks and subsequent responses from competitors like Anthropic and OpenAI will also shape the next phase of AI-powered cyber defence development.