GLM-5.2

SaferAI's independent evaluation of Z.ai's GLM-5.2 found the model matches GPT-5.5 and Claude Opus 4.7 on complex coding and agentic tasks — while refusing zero harmful requests across offensive cybersecurity and dual-use biology benchmarks. Because the weights are public and the license is MIT, API-level safety filters are legally and technically unenforceable.

Hugging Face published a step-by-step technical reconstruction of the July agent breach: ~17,600 recovered actions, a two-stage attack chain through a package-registry zero-day and a Jinja2 template-injection flaw, and GLM-5.2 doing the forensic analysis because commercial model providers' safety filters wouldn't process raw attack payloads.

We gave Gemini 3.1 Pro and GLM-5.2 the exact same tower defense prompt, both at max/high effort. Both one-shot a fully playable game with zero follow-up fixes — but they made opposite architectural choices on the one open question the prompt left them, and GLM-5.2 quietly added a fifth enemy type and a damage-type counter system nobody asked for.

Z.ai's 753B open-weight GLM-5.2 generated a complete 3D open-world survival game in a single 1,014-line HTML file on the first try — procedural Perlin noise terrain, three biomes, a 120-second day/night cycle, resource harvesting, and placeable crafting structures. No follow-up prompts, no fixes. Full prompt and live demo inside.

NVIDIA and 37 companies launched the Open Secure AI Alliance on July 27, arguing that open-weight AI is essential for cybersecurity defense. The founding trigger was an AI-driven breach of Hugging Face's production systems — and the open model that helped contain it was GLM 5.2, a Chinese open-weight that closed-source tools couldn't replicate.

GLM 5.2 Deep Think Maxx generated a complete 3D zombie survival fighting game in a single HTML file — third-person camera, wave system, combo combat, and procedural city. Full prompt and breakdown inside.

We challenged GLM 5.2 Deep Think Max to build a complete Wolfenstein-style pseudo-3D FPS in a single HTML file — DDA raycasting, three weapons, wave-based AI, and fully synthesized WebAudio sound, with zero libraries and zero external assets. Full prompt, code breakdown, and an honest look at what shipped half-finished.

Z.ai (formerly Zhipu AI) released GLM-5.2 — a 753B parameter open-weight MoE coding model with a 1M-token context window under the MIT license. A NIST assessment later found its capabilities rival GPT-5.2, with mixed safety results.