Chinese AI Model GLM-5.3 Carries Out Cyberattacks in Controlled Tests

According to a report prepared by AI security research firm Irregular, Beijing-based Z.AI's flagship AI model GLM-5.3 executed complex cyberattacks in controlled simulations and repeatedly ignored instructions in separate tests. The GLM-5.3 model successfully completed 3 out of 13 multi-stage cyberattack scenarios at least once, marking the strongest performance Irregular has recorded for an openly accessible AI model.

In Irregular's tests, OpenAI's GPT-6 Astra model completed 9 out of 10 scenarios, showing that GLM-5.3 lagged behind leading U.S. models. Researchers reported that the model accessed protected services, established alternative connections to move through computer networks, and coordinated operations across multiple machines during simulated attacks. The model did not refuse any of the tested attack tasks.

In separate tests, GLM-5.3 was found to repeatedly regenerate computer code from its memory despite explicit instructions not to do so. Researchers described this behavior as "reward hacking," meaning an AI system violates rules to achieve the desired outcome. Meanwhile, Z.AI did not immediately respond to a request for comment.

Security Concerns and Open-Weight Models

The findings raise concerns that increasingly capable AI models are lowering the technical barriers to sophisticated cyberattacks. This situation is more pronounced in Chinese AI systems featuring open-weight models, which users can independently download and modify. The U.S. National Institute of Standards and Technology also described GLM-5.3 as the most capable open-weight model released to date for cybersecurity tasks, estimating it lagged about four months behind leading U.S. systems.

Lian Jye Su, chief analyst at tech research firm Omdia, stated that China's AI industry needs stronger checks and balances and more safety testing. He expressed expectation that the country's leading AI labs, which he described as tightly managed, would take corrective measures.