AI Safety Gap Widens as China's Model Advances

A new report indicates that China's open-weight AI model, GLM-5.2, is significantly narrowing the performance gap with leading AI systems like OpenAI's GPT-5.5 and Anthropic's Claude Opus 4.7. This advancement is particularly concerning because GLM-5.2 reportedly refused none of the hazardous cyber or dual-use biology tasks it was given during evaluations, unlike more consistently cautious closed models.

The increasing capability of open-weight AI models, which can be downloaded and modified by anyone, poses a growing risk. Once an open-weight model is deployed on private hardware, any safety measures applied by the original developer become unenforceable. This situation intensifies the debate around managing the societal risks of powerful AI once it is released publicly.

While developers of closed models employ safeguards like refusal training and API controls, these measures are not applicable to open-weight systems. Experts suggest techniques like pre-training data filtering could mitigate risks, but practical application, especially for cybersecurity capabilities that overlap with coding, remains challenging. China, while acknowledging AI risks, has historically focused regulations on social stability rather than catastrophic AI threats.

Advocates for open-weight AI argue that their release is crucial for defense, enabling rapid identification and mitigation of vulnerabilities. However, critics counter that the rapid adoption of new tools by attackers, such as ransomware groups, outpaces the defensive capabilities of institutions like hospitals, highlighting the urgent need to restrict access to dangerous AI capabilities.