GLM-5.3 Sets a New Benchmark in Cybersecurity
A serious vulnerability in Cursor has been identified.
1Launch of GLM-5.3
Chinese AI startup Z.ai announced the release of GLM-5.3, the latest model in the GLM series. This model offers significant advancements in long-horizon coding and cybersecurity. GLM-5.3 stands out particularly for its cybersecurity capabilities, which have developed faster than expected. The model is currently available only through GLM Coding Plan and ZCode environments.
2Vulnerability in Cursor
GLM-5.3 identified a serious vulnerability in Cursor, an AI coding startup recently acquired by SpaceX. Z.ai developer advocate Lou shared this finding on X. VentureBeat is awaiting confirmation from Cursor. This discovery highlights GLM-5.3's impressive cybersecurity capabilities.
3Post-Training Scaling
GLM-5.3 uses the same base model as GLM-5.2 but has achieved significant improvements through post-training scaling. This scaling was achieved through more environments, diverse tasks, and additional reinforcement learning computations. Z.ai is testing how far a frontier-scale base model can be pushed without another expensive pretraining cycle.
4Coding Performance
GLM-5.3 improved from 4.6 to 28.3 on Terminal-Bench 3.0, from 46.2 to 66.9 on DeepSWE v1.1, and from 26.2 to 48.2 on AutomationBench. These improvements demonstrate the model's coding capabilities. However, GLM-5.3 does not lead in every area; it trails behind competitors like GPT-5.6 Sol and Claude Fable 5 in some benchmarks.
5Cybersecurity Advancements
GLM-5.3's cybersecurity capabilities developed faster than expected. The model scored 84.5% on CyberGym, surpassing GLM-5.2's 77.2%. However, it scored 54.4% on ExploitBench, lagging behind GPT-5.6 Sol and Mythos 5. This advancement shows the model's growing capability in security research.
6Security Findings
Z.ai, in collaboration with security teams in China, identified 2,436 vulnerabilities across 269 projects. Of these, 1,097 are critical or high severity. Currently, 53 vulnerabilities have been publicly disclosed, while 2,383 remain embargoed. This highlights the model's dual utility in software engineering and security research.
7Operational Efficiency
GLM-5.3 achieved a 34.5% score on Z.ai's private Code Bench while consuming approximately 75,000 output tokens per task. This is a significant improvement over GLM-5.2's 23.4% score and 96,000 token consumption. This efficiency reduces operational costs for enterprises deploying coding agents.
8Advanced Controls
Z.ai introduced a 'trusted access' approach for the model's more advanced capabilities. This ensures controlled use of sensitive functionalities. Reuters reported Z.ai's introduction of these controls. This step aims to prevent misuse of the model.
9Future Implications
GLM-5.3's advancements in cybersecurity and coding could expand the future applications of AI models. However, the risk of misuse also increases. Z.ai's work in this area demonstrates how to manage a technology that is both beneficial and potentially dangerous.
Original article: VentureBeat — GLM-5.3 Siber Güvenlikte Yeni Bir Çığır Açtı