An AI model just found a serious security flaw in the software that developers use to write AI-assisted code.
Z.ai released GLM-5.3 on August 14, 2026, and the model’s headline moment came quickly: it identified a significant vulnerability in Cursor, the AI-powered code editor. The discovery was flagged by security researcher Joshua Saxe.
The benchmark numbers tell a story
On the CyberGym vulnerability discovery benchmark, the model scored 84.5%, edging out Mythos 5 at 83.8% and GPT-5.6 Sol at 83.6%.
The ExploitBench results are where things get genuinely striking. GLM-5.3 scored 54.4% on that benchmark, which measures a model’s ability to reason through and execute exploits. Its predecessor, GLM-5.2, scored roughly half that.
Z.ai attributes the entire improvement to post-training reinforcement learning in security-focused environments. The base model itself was not changed.
2,436 vulnerabilities and counting
Since Z.ai introduced GLM-5.2, its models have collectively identified 2,436 vulnerabilities across 269 projects. Of those, 1,097 are classified as critical or high severity.
The scope of what these models have found spans kernels, operating systems, browsers, and protocols. Some of the vulnerabilities discovered had gone undetected since 1981.
GLM-5.3 also offers users flexibility in how much cognitive effort the model applies to a given problem, with settings for low, high, or maximum thinking effort. The feature cannot be disabled entirely.
Open weights, delayed by safety review
Z.ai released GLM-5.3 as an open-weights model, meaning external researchers and developers can download and run it without routing everything through a proprietary API.
The open-weights release was delayed by two weeks to allow for a safety evaluation. A model that can identify and reason about exploits at this performance level is, almost by definition, a dual-use tool.