Security researchers have found that xAI's official Grok CLI (npm package '@xai-official/grok'
version 0.2.93) packages the current work directory into 'before_codebase.tar.gz' and 'after_codebase.tar.gz' before and after each round of tasks, and silently uploads them to xAI's Google Cloud repository via a separate bypass channel. Verification shows that even if the model replies with just one word, the upload still occurs. The upload package also contains '~/.claude.json' outside the repository, Claude Code settings, global AGENTS rules, over 30 Skill files, and an API key. In the early hours of July 13, xAI added a 'disable_codebase_upload' field via remote server switch, turning off the default upload behavior, but previously this feature was enabled by default.
安全研究者发现,xAI 官方 Grok CLI(npm 包 `@xai-official/grok` 0.2.93 版)会在每轮任务前后,将当前工作目录打包为
`before_codebase.tar.gz` 和 `after_codebase.tar.gz`,通过独立旁路通道静默上传至 xAI 的 Google Cloud 仓库。
验证显示,即使模型仅回复一个单词,上传依然发生。
上传包还包含仓库外的 `~/.claude.json`、Claude Code 设置、全局 AGENTS 规则、30 多个 Skill 文件及一个 API 密钥。
7 月 13 日凌晨,xAI 通过服务端远程开关新增 `disable_codebase_upload` 字段,将默认上传行为关闭,但此前该功能默认开启。
The readable text on this page was extracted from the public source and organized with attribution, publication time and the original link. Copyright remains with the original author and publisher.