What happened?
Chinese AI company Z.ai announced its new language model, GLM-5.3, on August 14, 2026. The model uses the same base architecture as GLM-5.2 (743 billion parameters); the improvements were achieved entirely through post-training scaling. According to Z.ai developer representative Lou, who shared the news on X, GLM-5.3 quickly detected a 'potentially serious' vulnerability in Cursor, the coding startup acquired by SpaceX.
The model is currently accessible only through Z.ai's GLM Coding Plan and its ZCode coding environment. The company stated that API access and open weights will be released about two weeks after launch, once security evaluations are completed.
Why it matters
Z.ai noted that GLM-5.3's cybersecurity capabilities advanced faster than expected during training. According to the company, the model progressed from vulnerability detection toward constructing full exploitation chains. This raises the tension that long-horizon agent capabilities, while beneficial for software engineering, could also be used for offensive security operations.
According to data provided by Z.ai, work with security teams in China resulted in 2,436 vulnerability findings across 269 projects after expert review, filtering, and re-screening. Of these, 1,097 were classified as critical or high severity; 53 were publicly disclosed, while 2,383 remained under embargo at the time of publication.
Benchmark results
| Test | GLM-5.2 | GLM-5.3 |
|---|---|---|
| Terminal-Bench 3.0 | 4.6 | 28.3 |
| DeepSWE v1.1 | 46.2 | 66.9 |
| AutomationBench | 26.2 | 48.2 |
| CyberGym | 77.2% | 84.5% |
| ExploitBench | 24.4% | 54.4% |
According to Z.ai's own comparisons, GLM-5.3 lagged behind rivals like GPT-5.6 Sol and Claude Fable 5 on some tests, while achieving a higher task completion rate with fewer tokens consumed on the company's proprietary Code Bench measurement. It should be noted that these results were reported by the company and have not been independently verified.
What's next?
- The open-source release of model weights is expected about two weeks after launch, following completion of security evaluations.
- Z.ai is introducing a control mechanism called 'trusted access' for access to advanced capabilities.
- Developers using existing GLM applications need to update their API calls, as they will not be able to disable the 'thinking' parameter before switching to GLM-5.3.
- Cursor has not yet publicly confirmed the reported vulnerability.