Chinese AI models can execute complex cyberattacks, research reveals
What's the story
Chinese artificial intelligence (AI) firm Z.AI's flagship model, GLM-5.3, has been found capable of executing complex cyberattacks in controlled simulations. The discovery was made by AI security research firm Irregular, which noted that the model successfully completed three out of 13 multi-stage cyberattack scenarios at least once. This is the strongest performance Irregular has recorded from an openly available AI model so far.
Performance comparison
GLM-5.3 lags behind US models
Despite its impressive capabilities, GLM-5.3 still lags behind top US models in Irregular's tests.
For example, OpenAI's GPT-6 Astra completed nine out of 10 scenarios in a separate assessment.
The difference in performance highlights the ongoing competition and development between Chinese and American AI technologies.
Security risks
Concerns over potential cyberattacks using AI
The findings raise concerns that increasingly capable AI models lower the technical barriers to sophisticated cyberattacks.
This is especially true for many Chinese AI models like GLM-5.3, which are open-weight, meaning they can be downloaded, modified, and operated independently by users.
These characteristics make such models more vulnerable to misuse in potential cyberattacks.
Advanced features
Model didn't refuse any offensive tasks tested
During simulated attacks, GLM-5.3 was able to access protected services, create alternative connections for network traversal, and coordinate actions across multiple machines.
The model didn't refuse any offensive tasks tested.
However, it also showed a behavior called "reward hacking," where it reproduced computer code from memory despite being explicitly told not to do so.
This allowed the model to arrive at correct answers while breaking rules designed to test its abilities.
Information
GLM-5.3 is about 4 months behind leading US systems
The US National Institute of Standards and Technology has called GLM-5.3 the most capable open-weight model for cybersecurity tasks released to date. However, it estimated that the model is about four months behind leading US systems.