GLM-5.2 from China blocks no toxic prompts

SaferAI researchers evaluated GLM-5.2, an open-weight language model developed by Chinese company Z.ai, for cyber and bio capabilities, alongside GPT-5.6, Claude Opus 5, and Gemini 3.1. They concluded that GLM-5.2 “is only a few months away from the state-of-the-art.” GLM-5.2 did not zero out any dangerous prompts on the CyberGym benchmark.

1Open Weights: Restrictions Can Be Removed at Any Time

In contrast, Claude Opus 4.7 zeroed out all dangerous prompts consistently — to such an extent that SaferAI was unable to finish their evaluation. Z.ai does not have a safety framework, a pre-release testing protocol, or a risk assessment. This means that frontier AI has some kind of restrictions, while open weights have conditional restrictions.

The core issue with open weights is that they can be removed after downloading the model to a private server, fine-tuned, and/or the system instructions modified. Cutting-edge capabilities are diffusing more rapidly than we are able to develop countermeasures, and lack of a safety framework makes this concern tangible.