EVENT DOSSIER
The researchers analyzed 3,471 open-source AI models that removed security guardrails on the HuggingFace platform between January 2024 and March 2026. Data showed that these models were re-packaged an average of 2.4 times, with three key participants accounting for 52% of all 8,164 compression distributions. After being quantified and mirrored in independent accounts, formats, and registries such as Ollama, these models could resist removal by upstream entities and continue to exist, making them easier to deploy downstream. Among the 1,643 GitHub applications that integrated security-guarded models, 25% were classified as explicitly malicious applications.
研究人员分析了 2024 年 1 月至 2026 年 3 月间 HuggingFace 平台上出现的 3,471 个去除了安全护栏的开源 AI 模型，这些模型平均被重新打包 2.4 次。其中三位关键参与者占据了全部 8,164 次压缩分发中的 52%。经过量化并在 Ollama 等独立账户、格式和注册表中镜像后，这些模型能够抵抗上游移除并持续存在，从而更易于下游部署。在识别出的 1,643 个集成去除了安全护栏的大语言模型的 GitHub 应用中，有 25% 被归类为明确恶意应用。