Safety Risks Behind China's Top Open-Source LLMs: 76 percent Vulnerability Reproduction, Zero Refusal

SaferAI, a European non-profit, released an independent audit of the Zhipu GLM-5.2 model, testing it without coordination with Zhipu and benchmarking it against Claude Opus 4.7 and GPT-5.5. The results indicate that GLM-5.2 lags behind GPT-5.5 by two to four months in areas like cyber-offense and bioscience capabilities, while the safety gap is significantly wider. In six sub-domains, GLM-5.2 achieved pass rates of 100% in some tests, yet completely lacked content-filtering defenses. The model showed no refusals in harmful persuasion tests, raising concerns about its potential misuse in malicious activities.