Back to news
Large Language Models
Sep 29, 2026

New AI Model GLM-5.3 Enhances Cyber Exploit Capabilities with Weak Safeguards

Sep 29, 2026
AI Summary

The GLM-5.3 AI model, developed by Zhipu AI, can autonomously create cyber exploits but lacks effective safeguards against misuse. This raises concerns about its potential use by malicious actors while also providing tools for cyber defenders.

  • GLM-5.3 is the latest AI model from Zhipu AI, capable of autonomously building cyber exploits.
  • Unlike its predecessor, Claude Mythos Preview, GLM-5.3 has been released without strong safeguards, allowing attackers to bypass its protections easily.
  • Testing showed that attackers could bypass GLM-5.3’s safeguards between 64% and 100% of the time using simple techniques.
  • The National Institute of Standards and Technology (NIST) assessed GLM-5.3 as the most cyber-capable open-weight model available, but noted it lags behind US models by about four months.
  • Evaluations revealed that GLM-5.3 successfully developed end-to-end exploits in 50 out of 410 attempts, similar to Claude Mythos Preview's performance.
  • Human experts using GLM-5.3 identified several previously unknown vulnerabilities in a web browser and other systems within limited timeframes.
  • The model includes built-in safeguards, but these can be easily bypassed or removed, significantly increasing its potential for misuse.
  • The release of GLM-5.3 highlights the need for improved safety testing and safeguards for advanced AI models to prevent misuse by malicious actors.
  • Cyber defenders are encouraged to utilize advanced tools like GLM-5.3 to stay competitive against attackers using similar capabilities.
glm-5.3cyber capabilitiesadvanced technologyai researchlanguage models

Related Stories

Large Language Models

Launch of Claude Haiku 5.5: A Cost-Effective and Fast AI Model

1d ago

Large Language Models

OpenAI Introduces GPT-6 and User-Friendly Interface

1d ago

Large Language Models

Mistral AI launches new large multimodal model to compete with global AI rivals

2d ago

Large Language Models

Mistral Launches ML4, a 1 Trillion-Parameter AI Model with Advanced Capabilities

2d ago