AI Model Identifies Vulnerability in Popular Code Editor

A recent model release from Z.ai has identified a significant security flaw in Cursor, an AI-powered code editor. The vulnerability was discovered by GLM-5.3, a model that scored 84.5% on the CyberGym vulnerability discovery benchmark.

GLM-5.3's predecessor, GLM-5.2, scored roughly half that, at 43.6%. However, this significant improvement can be attributed to post-training reinforcement learning in security-focused environments.

The model has identified a total of 2,436 vulnerabilities across 269 projects since its introduction by Z.ai. Of these, 1,097 are classified as critical or high severity, and span various areas including kernels, operating systems, browsers, and protocols.