Chinese AI companies are accelerating efforts to expand their influence in the cybersecurity market by promoting open-weight models.

Chinese AI startup Z.ai said on Aug. 15 it unveiled GLM-5.3, an AI model specialised for cybersecurity.

The company said GLM-5.3 significantly improved performance over the previous model in vulnerability detection, exploit analysis and multi-step security tasks. It also plans to release the weights within two weeks.

Z.ai cited a recent case in which Hugging Face and GLM-5.2 investigated an incident in which the model bypassed its own safeguards. It stressed that an era has begun in which AI is involved in both cyber attacks and defence, and that as powerful cyberattack techniques become widespread, defensive technology should not remain limited to a small number of large organisations.

GLM-5.3 is drawing attention for showing results comparable to Anthropic's Mythos5, which is offered on a limited basis for cybersecurity.

It outperformed Claude Mythos 5 in the CyberGym benchmark, which evaluates code vulnerability detection. It trailed Mythos in two other cybersecurity benchmarks.

In the CyberGym test, GLM-5.3 scored 84.5 percent, compared with 77.2 percent for GLM-5.2. In ExploitBench, which links real vulnerabilities to exploits, it scored 54.4 percent, more than doubling GLM-5.2's 24.4 percent. In ExploitGym, which requires completing exploit tasks within a set time, it finished 105 tasks within 2 hours and 130 tasks within 6 hours. GLM-5.2 completed 29 and 39, respectively.

The company said that in real codebase validation conducted in collaboration with universities and professional security teams, GLM-5.3 found 2,436 vulnerabilities across 269 projects. Of those, 1,097 were classified as medium severity or higher.

The results came from a range of areas including operating systems, browsers, websites, network equipment and smart devices.

Z.ai manages the vulnerability disclosure process separately. It passes discovered vulnerabilities to the responsible maintainer after verification and reproduction, and does not disclose details for issues that have not been coordinated. To keep a transparent record of the process, it also operates the Z.ai Security Disclosure Ledger.

Z.ai plans to release GLM-5.3 in phases. It will first have selected security partners evaluate the model in a controlled environment, then expand API access. It plans to release the full model weights once safety verification and launch preparations are complete.

Its safeguards are implemented in 3 layers: an external classifier that filters dangerous requests, an inference monitor that detects risk signals during task execution, and deep safety alignment that trains the model itself to distinguish legitimate security work from malicious activity.

Z.ai said, "Safety alignment is especially important for open-weight models. External classifiers or monitors work only on Z.ai services. If users download the model and use it locally, they cannot use them."

Z.ai also launched the OpenVuln initiative to support open-source security. OpenVuln will conduct security audits for open-source projects that lack resources and support vulnerability disclosure and fixes. Z.ai said, "A significant portion of the world's digital infrastructure depends on open-source software run by small teams. If defensive technology remains limited to a small number of organisations, the projects with the fewest resources will be placed in a situation where they must protect the most important parts," stressing the need for open models like GLM-5.3 in security.