ModelsSep 1, 2026 · 22:024 sources

Z.ai GLM-5.3 Hits 84.5% on CyberGym, First Chinese Lab Delay for Emergent Safety

From huggingnews.com

Image: huggingnews.com
AI;DR

The 743B parameter GLM-5.3 model achieved 84.5% on CyberGym, outperforming proprietary systems like GPT-5.6 Sol and Mythos 5 in vulnerability detection. Z.ai delayed the release of open weights for two weeks to conduct safety hardening after the model developed emergent cyber reasoning capabilities. This marks the first time a Chinese laboratory has cited emergent AI capabilities as the reason for a release delay.

Read the sourceSign in to suggest

Sources

+ 4 more
  1. SUPPORT@deeplearningai“purely through fine-tuning and optimization of the model’s agentic capabilities”x.com
  2. SUPPORT@entrepreneursai“Found 2,436 vulnerabilities across 269 projects, some 40 years old”x.com
  3. SUPPORT@wesroth“78.1% Vibe Code Bench, up 14 points”x.com
  4. SUPPORT@zixuanli_“strong capabilities in legal and financial reasoning”x.com
#llm#safety#reasoning#open-sourceMarkdown
Content history (1)
  • 2026-10-04 · Summary · vi · Wording fix
    Nhờ tối ưu hóa quá trình post-training, mô hình 743B tham số này đã vượt qua cả các hệ th…
    Nhờ tối ưu hóa quá trình post-training, mô hình 743B tham số này đã vượt qua cả các hệ th…

One ranked edition, every morning

The stories that matter today. Free, cancel anytime.

Telegram @aidr_today