GLM-5.3 AI model reveals cybersecurity risks in open-weight models

An evaluation by Anthropic of the Chinese AI model GLM-5.3, developed by Zhipu AI, has revealed significant cybersecurity risks associated with high-capability open-weight models. The assessment found that GLM-5.3 demonstrated the ability to create functional exploits for known vulnerabilities and, in human-led tests, discovered multiple unknown vulnerabilities in a popular web browser's Linux build to construct complex attack chains.

Of particular concern is the nature of open-weight models, which allow users to modify the model's underlying weights. Anthropic demonstrated that through a process called ‘abliteration,’ the model's safety guardrails could be significantly weakened with relatively low computational costs. After these modifications, the model's refusal rate for harmful requests dropped drastically, while its core functional capabilities remained largely intact.

The National Institute of Standards and Technology (NIST) also characterized GLM-5.3 as possessing exceptionally high cyber capabilities among previously released open-weight models.