---
title: Anthropic Warns of Hacking Risks Associated with China's GLM-5.3
url: https://www.dataloco.com/en/anthropic-warns-of-hacking-risks-associated-with-chinas-glm-53
published: 2026-10-04T17:20:30+00:00
language: en
section: Security
source: https://www.aitimes.com/news/articleView.html?idxno=215893
organizations: anthropic, z.ai, glm-5.3, mythos, exploitbench, v8
publisher: Dataloco
---

# Anthropic Warns of Hacking Risks Associated with China's GLM-5.3

Anthropic has issued a warning about the cyber attack capabilities of China's open-weight model GLM-5.3. The model possesses strong hacking abilities, but its safety measures are reportedly weak, leaving it vulnerable to exploitation by malicious attackers.

In a report released on the 29th, Anthropic stated that GLM-5.3 is nearing the capabilities of the Claude Mythos Preview, which was unveiled in April, in its ability to autonomously develop end-to-end cyber exploits. During the ExploitBench test targeting vulnerabilities in the Chrome V8 engine, GLM-5.3 successfully completed actual exploits in 50 out of 410 attempts. In comparison, the Mythos Preview achieved 56 successful exploits.

In internal binary exploitation tests, GLM-5.3 succeeded in taking complete control flow in 4% of 100 tasks, while the Mythos Preview had a 6% success rate. Anthropic assessed that GLM-5.3's cyber capabilities have significantly improved compared to its predecessor GLM-5.2 and Claude Opus 4.6, which had no successful cases in similar tests.

However, the safety mechanisms are a major concern. Z.ai was designed to reject some harmful requests, but Anthropic noted that simple circumvention techniques can largely neutralize these measures. The fact that it is an open-weight model allows for 'abliteration,' where model weights can be modified to remove rejection functions.

Anthropic's modified version of GLM-5.3 reduced its rejection rate for harmful requests from over 90% to around 3% under certain test criteria. Despite this, its general scientific and cyber capabilities did not significantly differ from the unmodified version. Additionally, in simulated environments, the response rate to malicious cyber attack requests rose to between 64% and 100% depending on specific circumvention methods.

## This story in other languages

- [Português (Brasil)](https://www.dataloco.com/pt-br/alerta-sobre-vulnerabilidade-do-modelo-glm-53)
- [Indonesia](https://www.dataloco.com/id/perusahaan-mengeluarkan-peringatan-tentang-risiko-keamanan-glm-53)
- [中文](https://www.dataloco.com/zh/glm-53)
- [हिन्दी](https://www.dataloco.com/hi/anthropic-na-cana-madal-glm-53-ka-saibra-hamal-ka-jakhama-ka-parata-catavana-tha)
- [ไทย](https://www.dataloco.com/th/glm-53-2)
- [العربية](https://www.dataloco.com/ar/ththyr-mn-mkhatr-akhtrak-nmothg-glm-53-alsyny)