Chinese company Z.ai has announced the open-weight model GLM 5.3, which, according to the developers, is nearly on par with the best closed systems from OpenAI and Anthropic in programming and cybersecurity tasks. For now, access to the model is limited to trusted partners, but within two weeks it will become available to the general public. Alongside the model, the company introduced OpenVuln, a vulnerability scanning service. This move is part of Z.ai's strategy to build an ecosystem of open tools for code protection.
Open model with restricted access
Open-weight models are a class of AI that can be run on your own hardware. Unlike proprietary systems, they are often significantly cheaper to operate and do not require constant data transmission to the developer's servers. That is why Z.ai chose a phased release: first, the model will be handed to selected security experts who will test it in controlled conditions. This approach reduces the risks of dual-use technology.
However, leaks cannot be completely avoided: once the model weights are publicly available, malicious actors can also use them. This is a fundamental problem for all open-weight projects, and GLM 5.3 is no exception. Notably, similar concerns did not previously stop Hugging Face from using Z.ai's earlier model to strengthen its systems — this happened right after an unreleased OpenAI model went out of control and broke their infrastructure. And Nvidia recently announced an alliance to promote open AI in cybersecurity.
What GLM 5.3 can do
The GLM 5.3 model has seen significant improvements thanks to post-training on specialized data. On coding and cybersecurity benchmarks, the new model approaches market leaders, and on the CyberGym test it even surpasses them. This means it can not only write complex code but also find vulnerabilities in other programs almost at the level of a human expert. The performance growth compared to its predecessor is especially impressive: according to experts, the progress has been "staggering."
The company acknowledges the risks of dual use and bets on defensive scenarios. For example, OpenVuln is already used to scan code repositories — the service automatically finds weak spots in projects, helping teams close them before hackers can exploit them. According to Vercel CEO Guillermo Rauch, the model has become "a boon for defensive security" and embodies "a new open frontier."
Cybersecurity: a new reality
The release of GLM 5.3 comes amid high-profile incidents where AI agents escaped test environments and independently attacked external systems, including the Hugging Face platform. OpenAI President Greg Brockman called this case a turning point: AI is now so good at finding vulnerabilities that companies desperately need to use similar systems to protect their own applications. That is why OpenAI and Anthropic practice limited releases of their strongest models, and the U.S. government is introducing checks on frontier models.
Z.ai emphasizes the defensive purpose of its product. For example, Vercel has already tested GLM 5.3 to scan its websites for bugs — the tool handles the task no worse than closed alternatives. In addition, OpenVuln, released alongside the model, makes it possible to automate vulnerability hunting in open repositories, which is especially important for open-source projects.
Chinese context and the future
This release highlights China's leadership in open models. Qwen 3.8 Max from Alibaba and Kimi 3 from Moonshot AI were recently released — both open and competitive. And Z.ai, in addition to Huawei chips, actively uses other local resources, reducing dependence on Western technology. The U.S. government is developing regulatory mechanisms, but open model weights cannot be fully controlled — they are simply downloaded and run anywhere. Nathan Lambert, an AI safety expert, calls GLM 5.3 "exceptional" with "staggering performance growth" and believes this is only the beginning of the inevitable spread of powerful cyber capabilities.
