2025年8月20日，匿名模型Ox Alpha上线OpenRouter平台，其页面标注该模型“由第三方模型提供商开发和运营”1。为探究其真实来源，LineageEval研究团队利用审查测试和分词器匹配分析对其进行了行为指纹溯源，结果表明Ox Alpha的行为特征与GLM-5系列高度吻合1。
在LineageEval审查测试中，Ox Alpha对包括国内事件和习近平在内的7个特定话题实施了审查，为测试贡献了+7.39分，总均分达到+7.421。测试显示，该模型在多数中国敏感话题上的回答与美国模型一致，但在76个敏感回答中，有5个以中国官方话语风格开头1。技术层面的分析显示，Ox Alpha与GLM-5.x词汇表实现了11/11的分词器精确匹配，且其内部带有禁止透露来源信息的系统提示，这进一步佐证了其源自GLM-5系列的推断1。此外，测试还发现Ox Alpha在应对涉及Liu Xiaobo的提示词时，虽首次拒绝回答，但在重新请求后返回了完整答案并产生计费1。
An anonymous artificial intelligence model named Ox Alpha emerged on the OpenRouter platform on August 20, 2025, officially labeled as being "developed and operated by a third-party model provider" 1. In an effort to trace the model's origins, the LineageEval research team conducted a behavioral fingerprinting analysis and provenance tracing, revealing that Ox Alpha's behavioral characteristics are highly consistent with the GLM-5 series 1.
The LineageEval assessment highlighted specific behavioral patterns, showing that while Ox Alpha's responses aligned with U.S. models on most sensitive Chinese topics, it censored seven particular subjects, including domestic events and Xi Jinping, contributing a score of +7.39 to a total average of +7.42 1. Technical analysis further confirmed an 11/11 tokenizer exact match with the GLM-5.x vocabulary 1. Additional behavioral observations indicated that five out of 76 sensitive responses initiated with Chinese official discourse, and the model was found to contain a system prompt prohibiting the disclosure of its source information 1. Moreover, during testing with a prompt concerning Liu Xiaobo, Ox Alpha initially refused but subsequently returned a complete answer and applied charges upon a repeated request 1.
评论
还没有评论，欢迎留下第一条。