China’s Z.ai Drops GLM-5.3 And Its Cybersecurity Gains Are Raising Eyebrows

Z.ai has launched GLM-5.3, claiming major advances in coding and cybersecurity without changing the underlying GLM-5.2 base model.

Chinese AI company Z.ai has launched GLM-5.3, an aggressive new upgrade to its open-weights AI lineup that the company says dramatically improves coding, long-horizon reasoning, and cybersecurity.

The biggest surprise is how Z.ai says it achieved those gains: without replacing the underlying base model. Instead, the company scaled up its post-training process, adding more environments, tasks and computing power to push GLM-5.2’s capabilities much further.

The model was released on August 14 and is initially available through Z.ai’s coding products and API ecosystem. The company says it will release the model weights publicly in roughly two weeks, after additional safety evaluation and hardening.

GLM-5.3 Claims a Huge Coding Jump

Z.ai is positioning GLM-5.3 as a major contender for AI coding and autonomous software engineering.

According to the company’s reported results, GLM-5.3 delivered a 50% improvement over GLM-5.2 on Z.ai’s internal Code Bench. The jump becomes even more dramatic on certain public evaluations focused on real-world terminal and software-engineering tasks.

On Terminal-Bench 3.0, GLM-5.3 scored 28.3, compared with just 4.6 for GLM-5.2. DeepSWE v1.1 also rose sharply, from 46.2 to 66.9.

Those numbers matter because the latest generation of AI coding models is increasingly judged not simply on whether they can write code, but on whether they can independently navigate repositories, troubleshoot problems, and complete complicated, multi-step engineering assignments.

Z.ai says it achieved the improvement by scaling systems including IndexShare for long-context processing, SAO for reinforcement learning on long-horizon tasks, and Slime for large-scale asynchronous training. The company also expanded the number and diversity of training environments.

The Cybersecurity Results Are Even More Striking

The headline-grabbing part of GLM-5.3 may ultimately be its cybersecurity performance.

Z.ai says the model’s cybersecurity capabilities improved faster than expected as it scaled post-training. Rather than simply identifying isolated vulnerabilities, the company says GLM-5.3 demonstrated stronger reasoning across multiple stages of an exploitation chain.

On CyberGym, GLM-5.3 reportedly scored 84.5%, up from 77.2% for GLM-5.2. Its ExploitBench score more than doubled, rising from 24.4% to 54.4%.

On ExploitGym, Z.ai reported that GLM-5.3 completed 105 tasks within a two-hour budget and 130 within six hours, compared with 29 and 39, respectively, for GLM-5.2. The company notes that those budgets are normalized for inference speed and therefore should not be interpreted simply as real-world elapsed time.

That capability has a clear dual-use dimension: the same AI that can uncover vulnerabilities could potentially be used to discover weaknesses more efficiently.

AI Finds Thousands of Vulnerabilities

Z.ai also says it tested its cybersecurity systems against real-world open-source projects with security teams in China.

After expert review, screening and deduplication, the company says the effort uncovered 2,436 vulnerabilities across 269 open-source projects, including 1,097 medium-to-high-severity issues.

The reported findings span areas including operating systems, system kernels, browser engines, web applications, open-source infrastructure and network protocols. Z.ai has created a Security Disclosure Ledger to track the vulnerabilities and their disclosure status.

The company says 53 findings had been publicly disclosed at the time of the announcement, while thousands remained under embargo as it notified affected projects.

The Open-Weights Release Could Be the Real Test

For developers, researchers and AI competitors, the most important GLM-5.3 moment may still be ahead.

Z.ai says it will release the model weights approximately two weeks after launch, following safety testing and hardening. Until independent researchers can inspect and test the model themselves, many of the headline performance figures remain company-reported.

That makes the upcoming open-weights release crucial. If independent testing confirms the coding and cybersecurity gains, GLM-5.3 could become one of the most significant open AI models for software development and security research in 2026.

And if the cybersecurity capabilities prove as powerful outside Z.ai’s testing environment as they appear on paper, the model could also intensify the debate over how openly increasingly capable AI systems should be distributed.