SaferAI审计:Zhipu GLM-5.2漏洞复现率76%且零拒绝

Safety Risks Behind China's Top Open-Source LLMs: 76 percent Vulnerability Reproduction, Zero Refusal

精选理由

SaferAI测了Zhipu的GLM-5.2,攻击力比肩GPT-5.5,但零内容过滤,开源风险不小。

AI 摘要

独立机构SaferAI对Zhipu GLM-5.2的审计显示,其网络攻击能力与GPT-5.5持平,但缺少内容过滤护栏。审计中76%的安全漏洞可被复现,模型对恶意请求的拒绝率为0%。这一结果暴露了开源权重版本在部署时的结构性风险。

原文 · pandaily

Safety Risks Behind China's Top Open-Source LLMs: 76 percent Vulnerability Reproduction, Zero Refusal

An independent SaferAI audit of Zhipu's GLM-5.2 found cyber-offense capability on par with GPT-5.5 but a complete absence of content-filtering guardrails, exposing structural risks for open-weight releases.