# GLM-5.3 发布：网络安全能力大幅跃升

- 来源：Chubby♨️ (@kimmonismus)
- 发布时间：2026-08-14 15:15
- AIHOT 分数：54
- AIHOT 链接：https://aihot.virxact.com/items/cmssmdkwd09c2roffx7lgv8ht
- 原文链接：https://x.com/kimmonismus/status/2088162566719639717

## AI 摘要

智谱发布 GLM-5.3，743B 基础模型与 GLM-5.2 相同，提升完全来自扩展后训练。其 Terminal-Bench 3.0 得分从 4.6 升至 28.3，CyberGym 从 77.2% 升至 84.5%，ExploitBench 从 24.4% 升至 54.4%。智谱称该模型能推理多阶段漏洞利用并构建完整攻击链，进展快于预期。

## 正文

WHAT: Zai just launched GLM-5.3, and its biggest leap may be in cybersecurity.

The 743B base model remains unchanged (!) from GLM-5.2. Zai says the gains come entirely from scaling post-training across more environments, diverse tasks and long-horizon workflows.

Its results:

• Terminal-Bench 3.0: 4.6 → 28.3
• CyberGym: 77.2% → 84.5%
• ExploitBench: 24.4% → 54.4%

The company says GLM-5.3 became capable of reasoning across multiple stages of exploitation and constructing complete attack chains,faster than its researchers expected.

After this release, I unfortunately expect the US government to regulate open source.

Crazy release!

### 引用推文

> Z.ai：Introducing GLM-5.3: Built to Code. Ready for Cyber Defense. - Top-tier coding and agentic capabilities, achieved through post-training on the 743B base model -...