GLM-5.3 scored 84.5% on the benchmark, slightly higher than Fable 5’s 83.8% and GPT-5.6 Sol’s 83.6%.
On ExploitBench, which is more a measure of a model’s offensive ability, the model trailed the frontiers with 54.5% versus 78% for Fable 5 and 76.5% for GPT Sol.
Z.ai, the lab behind the model, says they will release it as open source with open weights in about two weeks, according to Reuters, after first letting «selected security partners» evaluate it in «controlled settings,» Z.ai writes on security.
GLM-5.3 will also have in place training to «distinguish legitimate security work from high-risk offensive activity» on the model level.
Z.ai does admit that, as an open weights model, they can’t be sure. «No safety system can eliminate every dual-use risk. Once model weights are public, no developer can guarantee control over every downstream modification or use. Model-level safeguards can raise the barrier to abuse, but they cannot provide absolute control,» they write.
The model will be restricted on Z.ai’s servers and in API use, but open weights mean that anyone can download it and tinker with it. Acquiring the expertise to do so might be the hardest part.
In theory, you would only need to invest around $300,000 to set up a computer capable enough to run the full model locally, or you could rent the compute. That would put advanced cyber capabilities within reach of well-funded organizations, universities or nation-backed groups.
GLM-5.3 has already been put to work on real codebases in cooperation with several Chinese security teams, finding 2,436 vulnerabilities across 269 projects. Pursuant to this, they have launched cvd.z.ai to «maintain a public record of findings,» after they are safe to disclose.
Read more: Z.ai’s launch page, Z.ai on security. Writeups on Reuters, South China Morning post. Discussion on Hacker News, r/Singularity.