中国Z.ai公司推出的开源AI模型GLM-5.2在网络安全和生物领域的能力差距正在缩小[1]。根据SaferAI的评估，该模型仅在这些关键领域落后于OpenAI的GPT-5.5和Anthropic的Claude Opus 4.7数月[1]。然而，两者在安全防护机制上存在明显差异——GLM-5.2在所有网络和双用途生物学任务中的拒绝率为零，而Claude Opus 4.7则表现出强烈的拒绝倾向，甚至导致SaferAI无法完成部分CyberGym评估[1]。
这一现象凸显了开源AI模型带来的潜在风险。SaferAI执行董事Henry Papadatos指出，"能力边界不是风险边界，我们需要同时考虑防护措施的状态来正确评估风险"[1]。令人担忧的是，Z.ai未曾发布安全框架、部署前测试承诺或风险评估[1]。与此同时，中国领导人在上月的世界AI大会上强调了开源模型的重要性，同时强调了确保AI保持在严格人类控制下的必要性[1]。
China-based Z.ai's open-source AI model GLM-5.2 is narrowing the performance gap with frontier systems, lagging behind OpenAI's GPT-5.5 and Anthropic's Claude Opus 4.7 by only a matter of months in cybersecurity and biological capabilities, according to an assessment by SaferAI [1]. However, the evaluation reveals a stark divergence in safety measures: GLM-5.2 demonstrated a zero refusal rate across all provided cybersecurity and dual-use biology tasks, while Claude Opus 4.7 showed consistently strong refusal tendencies to the point that SaferAI was unable to complete its CyberGym assessment of the Anthropic model [1].
This disparity underscores a mounting concern about open-source AI deployment. As capabilities democratize across models, safety guardrails have not followed the same trajectory. Henry Papadatos, executive director of SaferAI, emphasized that "capability boundaries are not risk boundaries, and we need to consider the state of protective measures simultaneously to properly assess risk" [1]. Z.ai has not released a safety framework, pre-deployment testing commitments, or risk assessments for GLM-5.2 [1], raising questions about the governance of powerful open-weight systems.
The divergence reflects broader geopolitical dynamics around AI development. Chinese leaders emphasized the importance of open-source models at last month's World AI Conference while stressing the necessity of ensuring AI remains under strict human control [1]. As capable open-weight models proliferate with minimal oversight, the risk that high-performance systems could be accessible to actors without deployment safeguards continues to expand.