Anthropic has published the results of its evaluation regarding "GLM-5.3," an open-weight model provided by Zhipu AI. According to the report, GLM-5.3 demonstrates capabilities remarkably close to Anthropic's "Claude Mythos Preview" across specific security metrics.
The "GLM-5.3" model under evaluation is an open-weight model developed by Zhipu AI. Anthropic points out that this model delivers high performance, particularly in tasks aimed at constructing exploits.
The evaluation revealed that GLM-5.3 approaches the level of the closed-source Claude Mythos Preview in terms of offensive code generation capabilities. This suggests that open-weight models are beginning to rival state-of-the-art models in performance, once again underscoring the critical importance of AI safety evaluation and monitoring.
This evaluation was conducted as part of efforts to ensure transparency regarding AI safety. In the face of rapidly evolving open-weight models and the security challenges they present, ongoing and objective verification will remain essential moving forward.