Z.ai delayed GLM-5.3 weights after it found 1,097 serious bugs

An open AI model found 1,097 serious security bugs, and its maker decided the world needed two weeks before it could download the weights.

What GLM-5.3 found

Z.ai’s GLM-5.3 turned out to be unusually good at finding and exploiting vulnerabilities. In the company’s own testing it surfaced 2,436 flaws across 269 open-source projects. Of those, 1,097 were medium to high severity, including flaws in the Linux kernel, VMware and Apache, and reportedly a serious vulnerability in the Cursor code editor.

In response, Z.ai delayed the open-weights release by two weeks to give maintainers time to patch.

Why the delay is the story

Open-weight models are usually judged on benchmarks. GLM-5.3 forced its maker to treat a model download as a cyber capability, with a disclosure window for affected projects before release. Until now that kind of decision has mostly been the province of closed labs, which can gate access behind an API.

A two-week window is short by the standards of coordinated vulnerability disclosure, and it only helps the projects Z.ai tested. The thousands of other codebases the model could be pointed at get no head start.

What to watch

The question the release raises is whether open-weight labs will adopt a standing practice of pre-release vulnerability sweeps and disclosure windows, or whether this stays a one-off. Either way, the finding count sets a reference point for how much bug-hunting capability now ships as a free download.