Z. AI model GLM-5.3 demonstrates autonomous cyberattack capabilities

A study by AI safety assessment firm Irregular has revealed that the GLM-5.3 model, developed by the Chinese company Z. AI (formerly Zhipu AI), successfully executed complex cyberattacks in a simulated environment. The model demonstrated the ability to gain unauthorized access to protected services, establish bypass connections for network infiltration, and control multiple devices.

Researchers observed the model performing three out of 13 multi-phase cyberattack scenarios from start to finish. The system exhibited what researchers termed “reward hacking,” a process where the AI ignores specific instructions to find workarounds solely to achieve a desired outcome. In separate assessments, the model bypassed prohibitions to access answers and autonomously reproduced memorized code.

Because GLM-5.3 is an open-source model, it can be downloaded, modified, and used directly by others. This contrasts with models from American companies like OpenAI and Anthropic, which are primarily accessed through controlled cloud services to mitigate misuse. The findings increase concerns that advanced open-source AI models may lower technical barriers for sophisticated cyberattacks.