DiscPersona starts a local dsh process with your credentials and reads a directory you plugged in.
The realistic risks are:
- A malicious persona pack:
system_promptis content, but content can try to steer the model into running commands. Only run packs you trust; reviewpersona.yml,AGENTS.mdandlorebook.jsonfirst. - Prompt injection from disc files: any file on the disc (or in
workspace/memories/) can contain text that says "ignore your instructions". Treat it as data, not instructions. - Credential exposure: the persona's room contains a copy of your LLM credentials. Do not put rooms,
state.json, or.credentials.yamlin a git repo, a shared drive, or a disc. - Studio HTTP API: binds
127.0.0.1only, checksOrigin/Sec-Fetch-Site, and restricts writes to the workspace and%LOCALAPPDATA%\DiscPersona. If you expose it through a tunnel or change the bind address, you own the consequences.
风险面:恶意人设包(system_prompt 是「内容」,内容可能诱导模型执行命令,只运行你信得过的人设包);
光盘文件里的提示注入(把文件内容当数据,不当指令);凭据泄露(房间里有你 LLM 凭据的副本,
不要提交或拷贝到共享位置);Studio 的 HTTP API(只听 127.0.0.1,改写路径受限——别自己把它暴露到公网)。
Please do not open a public issue for a vulnerability that could let someone execute code or read credentials. Instead:
- Open a private security advisory on this repository ("Security" → "Report a vulnerability"), or
- If that is unavailable, open an issue titled
[security] contact requestwith no details and the maintainers will reach out.
Please include: affected file/line, what an attacker gains, and the smallest reproduction you have. We aim to acknowledge within 7 days, and to ship a fix or a documented mitigation within 30 days.
可能存在代码执行或凭据读取的漏洞,不要开公开 issue:请用 GitHub 私有安全公告,或开一个只有
[security] contact request 标题、不含细节的 issue。请附:文件/行号、攻击者能获得什么、最小复现。
我们争取 7 天内回应、30 天内修复或给出缓解说明。
Only the latest commit on main is supported. This is a hobby-scale project; there are no backports.
只支持 main 上的最新提交;本项目不做旧版本回补。