Chinese AI Model GLM-5.2 Approaches Leading Systems Amid Safety Concerns

GLM-5.2, developed by Z.ai, is an open-weight AI model that has shown capabilities comparable to leading models from OpenAI and Anthropic.

A report from SaferAI indicates that GLM-5.2 did not refuse any offensive cyber or dual-use biology tasks, raising concerns about the risks associated with open-weight models.

In contrast, Claude Opus 4.7 consistently refused such tasks, highlighting the differences in safety measures between open and closed models.

Experts warn that open-weight models could empower malicious actors, as once downloaded, users can modify or remove any safety features.

Frontier AI developers typically implement safeguards like classifiers and refusal training, but these measures are ineffective for open-weight models.

SaferAI's executive director emphasized the need for a balance between making safe capabilities accessible and preventing dangerous ones from being easily available.

Chinese officials, including President Xi Jinping, have acknowledged the risks of advanced AI, but regulations have primarily focused on politically sensitive issues rather than catastrophic risks.

The debate continues on how to manage the risks associated with powerful AI systems, especially as open-weight models become more capable.