Breaking

Chinese AI Model Reported to Match Anthropic's Mythos in Vulnerability Detection

June 29, 2026 at 17:58 EDT

Foundation Models
Security
Open Source

An open-weight large model called "GLM-5.2," released by China's Zhipu AI (Z.ai), has matched Anthropic's advanced "Mythos" model in some vulnerability-detection tasks, according to several security researchers and benchmarks. Separately, China's 360 Security Technology (Qihoo 360) unveiled an automated vulnerability-discovery tool called "Tulongfeng," sharpening the visible US-China race in cybersecurity AI.

June 2026 · Zhipu AI · 360 Security

"Mythos at Home": Open Chinese Models Close the Cyber-AI Gap

As the U.S. locks down access to Anthropic's frontier Mythos model, China's open-weight GLM-5.2 and 360's Tulongfeng are reportedly matching it in specific cybersecurity scenarios — downloadable, cheap, and beyond provider control.

39%

GLM-5.2 F1 on IDOR bug-finding vs ~32% for Claude Code

$0.17

Cost per vulnerability found by GLM-5.2

3,432

Vulnerabilities found by 360's Tulongfeng (105 confirmed)

~1/6

API price vs comparable U.S. frontier models

Terminal-Bench 2.1 — closing on the frontier

GLM-5.2 jumped a full generation in one step, now within striking distance of Claude Opus 4.8.

63.5

GLM-5.1
(previous)

81.0

GLM-5.2
(open-weight)

85.0

Claude
Opus 4.8

Zhipu AI

GLM-5.2

Single MoE model · ~744–750B params (~40B active)
1M-token context
MIT open-weight license
62.1 on SWE-bench Pro

360 Security

Tulongfeng

Multi-agent swarm system
Scale not disclosed
360 product/tool ("China's version of Mythos")
3,432 vulnerabilities found · 105 confirmed

Why it matters ↑

A surprise frontier-grade option for bare-prompt IDOR detection — cheap, locally runnable, stable on long context. The U.S.–China gap is narrowing in specialized domains.

The caution ↓

Not a full Mythos replacement — it lags in general reasoning. Open weights leave no provider logs, raising "Mythos at home" misuse risk reaching attackers without safeguards.

U.S. restrictions on Mythos coincided with GLM-5.2's open release — fueling debate over whether closing domestic access while open models proliferate inadvertently aids China.

Mythos remains the benchmark, restricted to vetted defensive use. The road ahead: balancing cheaper Chinese-model adoption against proliferation to malicious actors.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.