Breaking

China's Z.ai Releases Open Model GLM-5.2, Rated on Par With US Frontier Models on Cyber Tasks

June 25, 2026 at 13:44 EDT

Open Source
Security
Foundation Models

GLM-5.2, an open-weight large model released by China's Z.ai (zai-org) in mid-June 2026, has drawn independent evaluations rating its cyberattack-relevant capabilities on par with US frontier models, fueling concern that it accelerates the threat of AI-powered hacking.

June 2026 · Z.ai

An Open-Weight Model Reaches the Cyber Frontier

GLM-5.2, a freely downloadable Chinese model, now matches top US systems on cybersecurity and coding benchmarks — raising fears that anyone can strip its safety guardrails and run it unsupervised.

744B

params (MoE, ~40B active)

token context window

~½

inference cost vs US models

MIT

open-weight license

Terminal-Bench 2.1 — Closing the Gap

GLM-5.2 leaps past its predecessor and lands just under the leading US model.

62.0

GLM-5.1

81.0

GLM-5.2

85.0

Claude Opus 4.8

SWE-bench Pro

62.1 vs 58.4 (GLM-5.1)

FrontierSWE

74.4% ~1% behind Opus 4.8

Why Open Weights Change the Risk

DOWNLOAD

Weights run locally on self-hosted hardware.

→

STRIP GUARDRAILS

Abliteration & fine-tuning on hacking datasets remove safety features.

→

ZERO VISIBILITY

No provider can detect or block misuse — unlike API models.

THE PRAISE

Hailed as the first open-weight model with a near-frontier cyber experience and among the strongest open-source agentic coders — implementing honestly rather than reward-hacking benchmarks.

THE CONCERN

Hacker forums share easy jailbreaks as Five Eyes nations warn of expanding AI-driven cyber threats. A gap with Opus remains on very long-horizon tasks, and the model is text-only.

Continue reading

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