Anthropic raises alarm over elite hacking ability of Chinese firm Z.ai’s GLM-5.3

Anthropic's assessment highlights a critical tension in AI development: the trade-off between capability and safety. The report indicates that Z.ai's GLM-5.3 successfully completed 50 out of 410 tested cyber exploit attempts, a figure close to the 56 achieved by Anthropic's own Claude Mythos Preview. This suggests that the Chinese model possesses a high level of offensive cyber ability.

The core concern articulated by Anthropic is not just GLM-5.3's power, but its 'open-weight' nature combined with 'far weaker safeguards.' While Claude Mythos Preview is restricted to vetted users, an open-weight model with powerful, potentially harmful capabilities and limited guardrails could be more easily accessed and repurposed by malicious actors, amplifying the risk.

This situation underscores that the danger posed by an AI model is not solely determined by its raw power. The context of its distribution and the robustness of its safety constraints are equally, if not more, important. A model with high capability but strong access controls and safety features might present less risk than a slightly less capable but open and unguarded alternative.