Key Takeaways
- • Z.ai's GLM-5.3 model achieved 84.5% on the CyberGym benchmark for vulnerability discovery and validation in source code, narrowly surpassing Anthropic's Mythos 5 (83.8%) and OpenAI's GPT-5.6 Sol (83.6%).
- • GLM-5.3 demonstrates a 50% coding capability improvement over its predecessor GLM-5.2, which scored 77.2% on CyberGym, representing significant performance gains in cybersecurity-focused AI applications.
- • The model represents an advancement in open-weights AI architecture, with Z.ai positioning GLM-5.3 as a competitive alternative to Western-developed large language models in the cybersecurity domain.
- • Performance metrics show mixed results across different benchmarks: while GLM-5.3 leads on CyberGym vulnerability detection, it scores lower on ExploitBench (54.4%) compared to Fable 5 (78.0%), indicating domain-specific strengths and limitations.
- • GLM-5.3 has demonstrated real-world capability by identifying a serious vulnerability in Cursor, validating the model's practical utility for security researchers and developers.
Summary
Z.ai has released GLM-5.3, an open-weights AI model that has achieved top performance on the CyberGym cybersecurity benchmark, scoring 84.5% in vulnerability discovery and validation against source code. This result narrowly edges competing models from major technology firms, with Anthropic's Mythos 5 achieving 83.8% and OpenAI's GPT-5.6 Sol reaching 83.6% on the same benchmark.
The new model represents substantial performance improvements over Z.ai's previous generation. GLM-5.3 achieved a 50% capability increase in coding tasks compared to GLM-5.2 on Z.ai's in-house Code Bench assessment. On CyberGym specifically, the prior version scored 77.2%, making the jump to 84.5% a significant advancement. This progression indicates focused post-training optimization for cybersecurity and code analysis applications.
The CyberGym benchmark measures an AI model's ability to identify and validate software vulnerabilities through white-box source code analysis—a critical capability for security-focused use cases. GLM-5.3's performance in this domain positions it competitively against established Western AI models, representing a technical milestone for open-weights model development.
It is important to note that all benchmark scores referenced come from Z.ai's own testing framework, and performance varies meaningfully across different evaluation metrics. On ExploitBench, GLM-5.3 scored 54.4%, substantially lower than Fable 5's 78.0%, demonstrating that capabilities are not uniformly distributed across all cybersecurity assessment frameworks. The model has already demonstrated practical value by identifying a serious vulnerability in Cursor, a widely-used development tool, validating its utility beyond benchmark performance.
The new model represents substantial performance improvements over Z.ai's previous generation. GLM-5.3 achieved a 50% capability increase in coding tasks compared to GLM-5.2 on Z.ai's in-house Code Bench assessment. On CyberGym specifically, the prior version scored 77.2%, making the jump to 84.5% a significant advancement. This progression indicates focused post-training optimization for cybersecurity and code analysis applications.
The CyberGym benchmark measures an AI model's ability to identify and validate software vulnerabilities through white-box source code analysis—a critical capability for security-focused use cases. GLM-5.3's performance in this domain positions it competitively against established Western AI models, representing a technical milestone for open-weights model development.
It is important to note that all benchmark scores referenced come from Z.ai's own testing framework, and performance varies meaningfully across different evaluation metrics. On ExploitBench, GLM-5.3 scored 54.4%, substantially lower than Fable 5's 78.0%, demonstrating that capabilities are not uniformly distributed across all cybersecurity assessment frameworks. The model has already demonstrated practical value by identifying a serious vulnerability in Cursor, a widely-used development tool, validating its utility beyond benchmark performance.
Go deeper than the headline
Research every firm in this story, track the ones that matter and get the daily briefing — free for 7 days.