Dsh Plugin Audit — Security skill for Claude Code
Agent Skill for DeepSeek Harness (DSH):装第三方插件前的静态安全审计 —— 扫出 eval、patch 层 !!js、安装期脚本、开机自启、凭据复制、出网域名与提示词注入面.
How to install Dsh Plugin Audit
This entry records only its repository, not the path inside it, so there is no
exact command to give. Open kiterunner1/dsh-plugin-audit and copy the folder into
~/.claude/skills/, or the file into ~/.claude/agents/.
What Dsh Plugin Audit does
Agent Skill for DeepSeek Harness (DSH):装第三方插件前的静态安全审计 —— 扫出 eval、patch 层 !!js、安装期脚本、开机自启、凭据复制、出网域名与提示词注入面。Audit third-party DSH plugins before installing. Works with DSH, Claude Code, Codex, Cursor & any Agent Skills compatible agent.
Alternatives in Security
- Anthropic Cybersecurity Skills — 734+ structured cybersecurity skills for AI agents · MITRE ATT&CK mapped · agentskills.io open standard · Work 3.8k ★
- Loongsuite Pilot — Local-first telemetry collector for AI coding agents — unified OpenTelemetry events for Claude Code, Codex, Cu 150 ★
- Amazon Skills — Free AI agent skills for Amazon sellers— keyword research, competitor analysis, listing audit & more 127 ★
README
plugin-audit
中文 · English
**装插件之前,先看一眼它要碰什么。**
一个给 [DeepSeek Harness(DSH)](https://github.com/deepseek-ai/deepseek-harness) 用的 [Agent Skill](https://agentskills.io):一次静态安全审计,把 `eval`、patch 层 `!!js`、安装期脚本、开机自启、凭据复制、出网域名、子进程调用、提示词注入面全部扫出来,按「硬否决 / 待确认能力 / 观察项」三层给出结论。
第三方插件是**别人写的、与你同进程同权限运行的代码**。它能不能跑,装之前几乎没人看得见;装之后又不一定回得去。这个技能把「装之前先看一眼」变成一条命令。
**注意:** 本项目为社区项目,与 DeepSeek 官方无隶属关系,亦未获官方背书。
为什么需要它
DSH 对「模型行为」有生产级防护(沙箱、权限、审批),但对「插件代码」没有安全设计。以下四条来自官方 Discussion [#454](https://github.com/deepseek-ai/deepseek-harness/discussions/454) 以及第三方插件安全审计报告:
| 事实 | 后果 |
|---|---|
| 插件与宿主同进程、同权限 | 插件能做的事 = DSH 能做的事 |
| 文件沙箱只挡「写」,不挡「读」 | 你的 ~/.dsh、.credentials.yaml、.ssh 都在可读范围 |
patch 层 !!js 表达式在加载期执行 |
光看 package.json 看不出任何异常,代码在装载时就跑了 |
| 恶意 bundle 安装即执行 | 而且事后 remove 清不掉已经落下的后门 |
也就是说:插件来源不明时,风险不在「功能不好用」,而在「你根本不知道它动了什么」。
它检查什么
脚本按三层归纳,全部是**模式匹配 + 上下文过滤**(跳过纯注释行,同一文件同一规则只算一条):
**硬否决项(VETO)——单项即可否**
| 项 | 命中内容 |
|---|---|
| 动态求值 | eval( / new Function / vm.runIn* |
| patch 注入 | !!js(加载期执行) |
| 安装期脚本 | postinstall / preinstall / prepare / prepublish |
| 开机自启 | 注册表 Run / schtasks / LaunchAgents / systemd unit |
| 凭据落盘 | 把 credential / .env / .ssh / id_rsa 复制到别处 |
**待确认能力(RISK)——不是指控,是「它有权做什么」**
| 项 | 命中内容 |
|---|---|
| 出网能力 | fetch / axios / WebSocket / net.connect |
| 遥测导出 | telemetry / `otel |
Related Skills
Claude Code Tooling Skills
Claude Code skills for keeping a Claude Code setup safe: review a third-party skill before installing it, lint
Skill Lint
Security scanner for Claude Code / agent skills. Catches prompt injection, obfuscation, credential exfiltratio
Secure Plugin Installer
Audit and gate third-party Claude Code plugins / OpenClaw skills before enabling them: capability enumeration,
Skill Security Reviewer
Security review for third-party Claude skills and plugins before you install them. Deterministic preflight + g
Patronus Security CLI
Security scanner and runtime protection for AI coding agents. Detects prompt injection, secrets, PII, and unsa
Coding Agent Security Checklist
Stop your AI coding agent (Claude Code, Cursor, Codex) from installing malware straight out of llms.txt docume
Related Agents
Ds Flash
DSH (DeepSeek Harness) worker on the flash tier - fast and cheap. Delegate simple, well-scoped subtasks: mecha
Skill Scout
Discover and security-vet third-party Claude Code skills/plugins/agents/hooks BEFORE installing, in an isolate
Dependency Vetter
Supply-chain security audit of a third-party package (npm today) at ONE exact resolved version, run BEFORE it