---
format: "aidr-story-markdown/v1"
id: "ea5b2484a47aba64f67bd4d7187764584aff083bcd584c55fbef3de4420ef9fa"
canonical_url: "https://aidr.today/ea5b2484?lang=en"
title: "Zai Org GLM-5.3 Triggers Safety Hold After Fine Tuning Hits Open Weights Top Spot"
lang: "en"
requested_lang: "en"
available_langs: ["en","vi"]
translation_fallback: null
fallback_fields: []
published_at: "2026-08-31T16:50:30.000Z"
category: "Models"
topics: ["llm","open-source","fine-tuning","safety","benchmark"]
source_urls: ["https://huggingnews.com/ai/update-zai-org-glm-53-triggers-safety-hold-after-fine-tuning-hits-open-w-1e0b6534","https://x.com/arena/status/2094440382440611935","https://x.com/DeepLearningAI/status/2094421042785665237"]
summary: "The Artificial Analysis Intelligence Index placed GLM-5.3 in a tie for the highest ranking among open weights models with a score of 60. Zai Org achieved these intelligence gains by fine-tuning GLM-5.2 instead of training a new base architecture. This post-training process produced emergent cybersecurity capabilities, highlighted by an 84.5% score on CyberGym, prompting a temporary safety hold on the model weights' release to evaluate exploit generation risks. Zai Org also launched GLM-5.3-Flash, which entered the Agent Arena ranking 4th among open source models and 19th overall. The model carries a $0.12 median cost per task and provides a 15.3% improvement in confirmed success, positioning it between DeepSeek V4 (High) and GPT-5.6 Luna (xHigh) on the Pareto frontier."
---

# Zai Org GLM\-5\.3 Triggers Safety Hold After Fine Tuning Hits Open Weights Top Spot

> [Open the canonical story](<https://aidr.today/ea5b2484?lang=en>)

**Published:** 2026-08-31T16:50:30.000Z
**Category:** Models
**Topics:** llm, open\-source, fine\-tuning, safety, benchmark

## Summary

The Artificial Analysis Intelligence Index placed GLM\-5\.3 in a tie for the highest ranking among open weights models with a score of 60\. Zai Org achieved these intelligence gains by fine\-tuning GLM\-5\.2 instead of training a new base architecture\. This post\-training process produced emergent cybersecurity capabilities, highlighted by an 84\.5% score on CyberGym, prompting a temporary safety hold on the model weights' release to evaluate exploit generation risks\. Zai Org also launched GLM\-5\.3\-Flash, which entered the Agent Arena ranking 4th among open source models and 19th overall\. The model carries a $0\.12 median cost per task and provides a 15\.3% improvement in confirmed success, positioning it between DeepSeek V4 \(High\) and GPT\-5\.6 Luna \(xHigh\) on the Pareto frontier\.

## Sources

- [Story source](<https://huggingnews.com/ai/update-zai-org-glm-53-triggers-safety-hold-after-fine-tuning-hits-open-w-1e0b6534>)
- [Story source](<https://x.com/arena/status/2094440382440611935>)
- [Supporting source](<https://x.com/DeepLearningAI/status/2094421042785665237>)