一个名为 OX Alpha 的神秘大语言模型在 OpenRouter 上线并攀升排行榜 1。该模型的系统提示词要求其严格自称“ox-alpha”，由未公开组织开发 1。通过提示词注入提取系统提示词后，模型回答：“I'm GLM, a large language model made by Z.ai” 1。
此外，gzip-NCD 压缩分析也独立验证了其与 GLM-5.3 最为接近 1。该分析使用的参考语料包含 60 个提示词，由 GPT-5.5、Claude Opus 5、Gemini 3.7 Flash、Gemini 3.1 Pro Preview 和 GLM-5.3 回答，共 293 条参考文本 1。OX Alpha 回答了其中 13 个提示词及 1 个新提示词，共 14 次查询 1。在 k=3、k=5、k=7、k=9 测试中，GLM-5.3 分别赢得 7/14、7/14、6/14、7/14 的投票，Claude Opus 5 为一致的第二名 1。
A mysterious large language model named OX Alpha recently launched on the OpenRouter platform and rapidly ascended its rankings 1. The model's system prompt strictly instructed it to identify itself as "ox-alpha" and attribute its creation to an undisclosed organization 1. However, users employed prompt injection techniques to extract the system prompt and feed it back to the model, prompting it to reveal its true identity by stating, "I'm GLM, a large language model made by Z.ai" 1.
To independently verify this claim, a gzip-NCD compression analysis was conducted for model attribution 1. The analysis utilized a reference corpus comprising 293 texts generated from 60 prompts answered by GPT-5.5, Claude Opus 5, Gemini 3.7 Flash, Gemini 3.1 Pro Preview, and GLM-5.3 1. OX Alpha was evaluated on 13 of these original prompts along with one additional prompt, resulting in 14 total queries 1. Across tests configured with k=3, k=5, k=7, and k=9, GLM-5.3 emerged as the closest match, winning 7/14, 7/14, 6/14, and 7/14 votes, respectively 1. Claude Opus 5 consistently placed second in the attribution tests 1.
评论
还没有评论，欢迎留下第一条。