Please do not open a public issue. Report privately through GitHub's private vulnerability reporting and include affected versions or commits, reproduction steps and the impact you observed.
请不要公开提交 issue。通过 GitHub 私密漏洞报告 提交,并附上受影响的版本或 commit、复现步骤和实际影响。
Only the latest commit on main receives fixes. 只修复 main 分支的最新版本。
Jingmei reads untrusted group messages and feeds them to an LLM that can call tools, so these are in scope:
精魅读取不可信的群消息并交给可调用工具的 LLM,以下问题都在范围内:
- run_js escape — reaching host objects, the file system, the network or environment variables from
run_js(threat model: docs/architecture.md). / 从run_js拿到宿主对象、文件、网络或环境变量。 - Secret disclosure — bot tokens or API keys appearing in logs, replies, session files or error messages. / bot token、API key 出现在日志、回复、会话文件或错误信息里。
- Server-side request forgery — media or tool fetches reaching private or internal addresses. / 媒体或工具请求访问到内网地址。
- Memory and privacy bypass — collecting data after
/forget, recalling another space's members, or leaking member profiles into public replies. //forget后仍收集数据、跨群召回成员、在公开回复中泄露成员档案。 - Authorization bypass — non-admins running
/contextor/compact, or a persona answering outside its configured spaces. / 非管理员执行管理员命令,或角色在未配置的空间发言。
Out of scope: weaknesses in third-party providers (Discord, Telegram, model providers, TypeSafe, Fish Audio) and attacks that require access to the host account or its .env.
不在范围内:第三方服务本身的问题,以及需要先拿到宿主账号或 .env 的攻击。