Z.ai releases GLM-5.3 AI model with cybersecurity capabilities

Chinese AI startup Z.ai has released its new language model, GLM-5.3, which includes significant improvements in handling long coding sequences and advanced cybersecurity capabilities.

The model has reportedly already identified a "potentially serious vulnerability" in Cursor, a coding startup recently acquired by SpaceX. Z.ai developer advocate Lou shared this information on X, and the company is awaiting confirmation from Cursor.

GLM-5.3 will initially be available through Z.ai's own GLM Coding Plan and ZCode development environments. API access and open weights are slated for release later, following completion of safety evaluations and hardening. Z.ai anticipates releasing the open weights approximately two weeks after launch.

According to the company, GLM-5.3 is based on the same foundational model as its predecessor, GLM-5.2. The advancements were achieved by scaling post-training across more diverse environments and tasks. This approach tests how far a frontier model can be pushed without a new, costly pre-training cycle. The company observed that cybersecurity capabilities advanced faster than expected during the scaling of post-training, particularly as tasks evolved from vulnerability identification to constructing exploitation chains.

Z.ai also plans to implement new controls around the model's advanced functions, including a "trusted access" method for sensitive operations. The company has collaborated with security teams in China, leading to the discovery of thousands of vulnerabilities across various projects. This development highlights the increasing tension for frontier model providers: the same long-horizon agent capabilities that make models useful for software development can also make them capable security researchers or potential offensive actors.