Chinese Zhipu, also known as Z.ai, announced this Friday (14) the launch of the artificial intelligence model GLM-5.3. The company says the system surpassed Anthropic's Mythos 5 in a key cybersecurity test, as China seeks to reduce the West's advantage in AI defense.
In the CyberGym benchmark, which measures the ability to identify and validate security flaws from source code, GLM-5.3 achieved 84.5% success. The percentage surpassed the 83.8% of Mythos and the 83.6% of OpenAI's GPT-5.6 Sol, according to Zhipu.
Lower result in ExploitBench
In another test, the Chinese model's performance was lower. In ExploitBench, which evaluates how far systems advance along the vulnerability exploitation scale, GLM-5.3 scored 54.4%, while Mythos and GPT-5.6 Sol recorded 78% and 76.5%, respectively.
Real vulnerabilities
The company said it tested GLM-5.3 with security teams in China on real code bases. After expert review, the model identified 2,436 vulnerabilities across 269 projects; reviewers classified 1,097 as medium to high severity.
