返回
原文
Rohan Paul@rohanpaul_ai
52AI 编辑部评分，满分 100

智谱 GLM-5.3 在漏洞发现等网络防御测试中超越 Anthropic Mythos 5

2026-08-14 19:25· 3小时前
AI 导读

智谱发布 GLM-5.3，基于 743B 基座模型后训练，在漏洞发现等网络防御测试中超越 Anthropic Mythos 5。其 Terminal Bench 3.0 从 4.6 升至 28.3，DeepSWE 达 66.9；CyberGym 得分 84.5%，但 ExploitBench 仅 54.4%，落后于 Mythos 5 的 78.0%。

China's Z .ai;s new model GLM-5.3 beats Anthropic's Mythos 5 in cyber-defence tests, i.e. at vulnerability discovery.

Coding jumped hard: Terminal Bench 3.0 went 4.6 → 28.3 and DeepSWE 46.2 → 66.9.

On CyberGym, which tests finding and validating source-code flaws, GLM-5.3 reports 84.5% vs. 83.8% for Mythos 5 and 83.6% for GPT-5.6 Sol.

But on ExploitBench, which requires deeper exploit development, GLM-5.3 drops to 54.4%, vs. 78.0% for Mythos 5 and 76.5% for GPT-5.6 Sol.

The same gap appears on ExploitGym, where Z .ai reports 105 completed tasks in 2 hours and 130 in 6, against Mythos 5's 181 and 247.

Its training environments increasingly resemble multi-day engineering work, forcing the model to diagnose, edit, test, and recover across long task chains.

Z .ai plans to publish GLM-5.3's weights after a two-week safety review, while limiting its most sensitive cyber functions to verified users.

Z.aiIntroducing 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 -...
Anthropic模型发布编码

来源：Rohan Paul · x.com

导出 Markdown
同一事件 · 1
返回
AI 评分 52/100在 X 查看原推
导出 Markdown

智谱 GLM-5.3 在漏洞发现等网络防御测试中超越 Anthropic Mythos 5

Rohan Paul · @rohanpaul_ai · X·2026-08-14 19:25·3小时前
AI 导读

智谱发布 GLM-5.3，基于 743B 基座模型后训练，在漏洞发现等网络防御测试中超越 Anthropic Mythos 5。其 Terminal Bench 3.0 从 4.6 升至 28.3，DeepSWE 达 66.9；CyberGym 得分 84.5%，但 ExploitBench 仅 54.4%，落后于 Mythos 5 的 78.0%。

China's Z .ai;s new model GLM-5.3 beats Anthropic's Mythos 5 in cyber-defence tests, i.e. at vulnerability discovery.

Coding jumped hard: Terminal Bench 3.0 went 4.6 → 28.3 and DeepSWE 46.2 → 66.9.

On CyberGym, which tests finding and validating source-code flaws, GLM-5.3 reports 84.5% vs. 83.8% for Mythos 5 and 83.6% for GPT-5.6 Sol.

But on ExploitBench, which requires deeper exploit development, GLM-5.3 drops to 54.4%, vs. 78.0% for Mythos 5 and 76.5% for GPT-5.6 Sol.

The same gap appears on ExploitGym, where Z .ai reports 105 completed tasks in 2 hours and 130 in 6, against Mythos 5's 181 and 247.

Its training environments increasingly resemble multi-day engineering work, forcing the model to diagnose, edit, test, and recover across long task chains.

Z .ai plans to publish GLM-5.3's weights after a two-week safety review, while limiting its most sensitive cyber functions to verified users.

Z.aiIntroducing 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 -...

来源：Rohan Paul· x.com

Anthropic模型发布编码
同一事件 · 1