Breaking

Zhipu Releases GLM-5.2, Said to Rival Claude in Flaw Detection

June 29, 2026 at 01:27 EDT

Foundation Models
Open Source
Security

China's Zhipu AI (operated by Z.ai) released its open-weight large language model GLM-5.2 in June 2026, with independent evaluations reportedly showing software-vulnerability detection on par with Anthropic's Claude Mythos. The company began rolling out GLM-5.2 to GLM Coding Plan subscribers on June 13, and on June 16 published the weights under an MIT license on Hugging Face and ModelScope.

June 2026 · Zhipu AI (Z.ai)

An open-weight Chinese model just matched Claude on cyber tasks

GLM-5.2 — MIT-licensed, downloadable, and runnable locally — beat Anthropic's top Claude on authorization-bypass detection, raising sharp questions about whether U.S. AI export controls still hold.

~744B

Total parameters (MoE), ~40B active

Token context window

~$0.17

Cost per vulnerability found

IDOR vulnerability detection (F1, prompt-only)

The open model edged out the restricted frontier model — higher is better.

39%

GLM-5.2

open-weight

32%

Claude Code

Opus 4.6

Terminal-Bench 2.1

81.0 vs Claude Opus 4.8 ~85.0

SWE-bench Pro

62.1 vs GPT-5.5 58.6

What developers praise

Cost efficiency and coding precision
Privacy and customization of local execution
MIT license — anyone can fine-tune

Concerns raised

Lags Claude Opus on general reasoning
Censorship risk; safety guards easily bypassed
Distillation suspicions, unanswered

The strategic question

Claude's offensive-cyber frontier model sits behind U.S. export controls. An open-weight model that matches it — runnable in an unmonitored local environment — tests whether those controls still work.

Zhipu AI, a Tsinghua University spinout, has built the GLM series for years. Its evaluation is now a symbolic case in U.S.–China AI competition.

Continue reading

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