Breaking

Chinese Open-Weight Model GLM-5.2 Matches Top US Systems in Cybersecurity Benchmarks

July 17, 2026 at 09:17 EDT

Foundation Models
Security
Open Source

Zhipu AI's newly released open-weight model, GLM-5.2, has outperformed Anthropic's Claude Code in an independent vulnerability-detection benchmark, signaling that Chinese systems are closing the gap with leading US models in cybersecurity tasks. The result adds to mounting evidence that the performance edge long enjoyed by American frontier labs is narrowing, particularly in coding and agentic security work.

June 16, 2026 · Zhipu AI / Z.ai

Open-weight GLM-5.2 beats Claude Code on a cyber-vulnerability benchmark

A freely downloadable Chinese model topped Anthropic's Claude Code at spotting IDOR flaws — using only a bare prompt — the latest sign the US frontier lead is narrowing in coding and security work.

744B

Total parameters (MoE), ~40B active per token

Token context window for long-horizon agents

~1/6

API price vs comparable frontier models · MIT license, self-hostable

The headline result — IDOR detection (F1 score)

Bare prompt, no scaffolding · ~$0.17 per vulnerability found

39%

GLM-5.2

open-weight

28–37%

Claude Code

recent Opus

Coding benchmarks — climbing toward the frontier

Terminal-Bench 2.1

GLM-5.2 · 81.0

GLM-5.1 · 63.5

Claude Opus 4.8 · 85.0

SWE-bench Pro

GLM-5.2 · 62.1

GLM-5.1 · 58.4

Why it matters

First open-weight model recommended as frontier-grade for cyber work
Low cost + self-hosting suits repository-scale scanning
MIT license avoids vendor lock-in and keeps sensitive code in-house

The caveats

Headline win rests on a single task and dataset
Claude can lead on other flaw classes (e.g. SSRF)
Stronger reward-hacking tendency; heavy GPU needs for self-hosting
Dual-use: freely downloadable capability could aid attackers

Continue reading

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