Zhipu's GLM-5.3 AI model outperforms Anthropic's Mythos 5 in cybersecurity
What's the story
Chinese artificial intelligence (AI) firm Zhipu, also known as Z.ai, has launched its flagship GLM-5.3 model. The company claims that the new model has outperformed Anthropic's frontier Mythos 5 in a crucial cybersecurity test. The development comes as China is trying to keep up with Western advancements in AI defense technology.
Performance metrics
GLM-5.3 outperformed Mythos on CyberGym benchmark
Zhipu has revealed that its GLM-5.3 model scored an impressive 84.5% on CyberGym, a benchmark for identifying and validating security flaws from source code.
This score is higher than that of Anthropic's Mythos at 83.8% and OpenAI's GPT-5.6 Sol at 83.6%.
However, the Chinese model fell short on ExploitBench, which measures how far AI models climb the exploitation ladder, with a score of just 54.4%.
Vulnerability detection
Model tested against real-world codebases
Zhipu has tested the GLM-5.3 model with security teams in China against real-world codebases.
The tests revealed a total of 2,436 vulnerabilities across 269 projects after expert review.
Out of these, 1,097 were rated medium to high severity by the company.
This highlights the potential of Zhipu's new model in identifying security flaws in software systems effectively.
Future plans
GLM-5.3 to be made public in 2 weeks
Zhipu has announced its plans to release the GLM-5.3 model publicly in about two weeks.
The company intends to complete security assessments and strengthen safeguards before the launch.
However, it will only allow verified users access to its most sensitive cybersecurity functions through a "trusted access" program.
This move is aimed at ensuring that advanced cyber-defense tools are available to developers of open-source software and smaller security teams.