GLM-5.3 Got Much Better at Breaking Software, So Z.ai Is Holding the Weights Back
Z.ai's new coding model tops the open-weights field, and its security skills grew faster than the company expected. The download will follow two weeks late, after a safety review.
Z.ai released GLM-5.3 yesterday and did something unusual for a lab that built its reputation on quick, permissive downloads: it announced the model without publishing the weights. The company says they will follow in about two weeks, once safety evaluation and hardening are done. GLM-5.2 had gone up on Hugging Face under an MIT license within days of launch, and that was the whole selling point.
The reason for the pause is in the numbers. Z.ai added vulnerability discovery material to its post-training mix, the tuning phase that comes after the expensive base training, expecting the model to get better at spotting individual software flaws. Instead, it says the skill kept compounding and the model started planning complete attack chains. On CyberGym, a test of finding and confirming real vulnerabilities from source code, GLM-5.3 scores 84.5 percent against 77.2 for its predecessor. On ExploitBench it more than doubles, from 24.4 to 54.4 percent.
The coding results are the other headline. GLM-5.3 uses the same 743 billion parameter base as GLM-5.2, with every gain coming from more post-training on more varied work environments, some of them representing several days of an experienced engineer’s work. Terminal-Bench 3.0 moves from 4.6 to 28.3, DeepSWE from 46.2 to 66.9. All figures are vendor-reported and Z.ai still trails GPT-5.6 Sol and Claude Fable 5 on several public coding tests.
There is a constructive side to the security capability. Working with security teams in China, Z.ai says its models have found 2,436 vulnerabilities across 269 open-source projects since GLM-5.2, of which 1,097 are rated critical or high. The oldest, the company says, was introduced in 1981. Fifty-three are publicly disclosed with CVE numbers; the rest sit under embargo in a public disclosure ledger.
Worth keeping in mind: the same ability that finds decades-old bugs for patching is the ability an attacker wants. That tension is exactly why the release is staggered, and it lands in the same fortnight that OpenAI slowed its Astra model over cyber concerns. Two labs, two continents, same problem.
What this means for you: if you write code, GLM-5.3 is available now through Z.ai’s API and its coding plan, and one API change may bite you: thinking mode can no longer be switched off. If you were waiting to self-host, the calendar now says end of August rather than today. And if you use software at all, which is everyone, the interesting part is the ledger: models are becoming genuinely good at finding old flaws in the code everything runs on, and for the moment the finders are publishing rather than selling.
Sources
Source: https://z.ai/blog/glm-5.3
Grok 4.6 Shows Up in GitHub Copilot, Two Days After Launch
xAI's newest model is now selectable in Copilot across eight places developers work, from VS Code and Xcode to the command line. The speed of these hand-offs says something about how the market works now.
