GLM-5.3's open release has yet to produce major attacks despite Anthropic's warnings about its Mythos-level cyber risk, undercutting calls to ban open models

The narrative that open AI models inherently pose greater cyber risks than closed ones is being challenged by real-world events.

Anthropic's warning about GLM-5.3 posing significant cyber risks has not materialized with major attacks since its open release. The author argues this situation exposes flaws in the current discourse surrounding open-weight AI models. While Anthropic's technical research on GLM-5.3's capabilities is considered reasonable, the report allegedly failed to address broader questions like the implications of banning open models or why Chinese companies deem them safe for release. Publicly available information suggests that closed models, not open ones, have been the primary source of documented cyberattacks to date. The author posits that both open-weight models and closed model APIs, despite varying safeguards, may be equally susceptible to misuse, challenging the narrative that closed systems are inherently safer. This perspective suggests that banning open models could paradoxically increase long-term cyber risk and hinder AI competitiveness.

The author's analysis suggests that the perceived cyber risk from open-weight models like GLM-5.3 may be overblown, as evidenced by the lack of significant attacks post-release. This challenges the prevailing discourse, often led by frontier labs and national security entities, which advocates for stricter controls on open models. The argument is made that the focus on banning open models overlooks the potential for misuse of closed-model APIs and ignores the possibility that open models could be crucial for cyber defense in certain sensitive environments.

This debate highlights a critical trade-off between AI competitiveness and perceived security risks, with the author suggesting that overly restrictive policies on open models could cede advantages to other nations and increase long-term vulnerability. The piece calls for a more nuanced understanding of AI risk, one that acknowledges the complexities of global AI development and the practicalities faced by different international actors, rather than relying on a simplified dichotomy of "open equals dangerous, closed equals safe."