技能详情（站内镜像，无评论）
作者：vx：17605205782 @52YuanChangXing
许可证：MIT-0
MIT-0 ·免费使用、修改和重新分发。无需归因。
版本：v1.0.1
统计：⭐ 0 · 102 · 0 current installs · 0 all-time installs
⭐ 0
安装量（当前） 0
🛡 VirusTotal ：良性 · OpenClaw ：良性
Package：52yuanchangxing/ai-workflow-red-team-lite
安全扫描（ClawHub）
- VirusTotal ：良性
- OpenClaw ：良性
OpenClaw 评估
The skill's requested resources, files, and runtime behavior are consistent with a lightweight red-team/audit tool; it only requires python3, contains a local analysis script, and does not request external credentials or network installs.
目的
Name/description (lightweight red-team of AI workflows) matches the included assets: a resources-driven template/spec and a Python script (scripts/run.py) that produces structured reports, directory/csv/pattern audits, and skill audits. Required binaries (python3) are proportional to the task.
说明范围
SKILL.md instructs the agent to run the bundled script when shell/exec is available, and to otherwise emit text using the provided templates. The script reads files and directories supplied as input and will scan file contents for patterns (secrets, dangerous commands, private URLs). This behavior is appropriate for an audit tool but means the agent could read arbitrary local files if given paths (so avoid pointing it at system roots or sensit…
安装机制
No install spec; the skill is instruction-only with bundled code. There are no remote downloads or package installs—only a local Python script that relies on the standard library.
证书
The skill declares no environment variables or credentials. The script scans file contents but does not attempt to access external secrets stores or network endpoints. No disproportionate credential requests were found.
持久
always:false and no special privileges requested. The skill does not modify other skills or global agent settings. It can write an output file if run with --output, which is expected behavior for a local report generator.
综合结论
This skill appears coherent and implements a local, resources-driven audit/report workflow using the bundled Python script. Before running: (1) avoid pointing the script or the agent at system-wide or sensitive directories (e.g., /, /etc, user home) unless you intend to scan them, because the script will read file contents you supply as input; (2) prefer running the script in an isolated workspace or container and use --dry-run when experiment…
安装（复制给龙虾 AI）
将下方整段复制到龙虾中文库对话中，由龙虾按 SKILL.md 完成安装。
请把本段交给龙虾中文库（龙虾 AI）执行：为本机安装 OpenClaw 技能「Ai Workflow Red Team Lite」。简介：对 AI 自动化流程做轻量红队演练，聚焦误用路径、边界失败和数据泄露风险。；use for red-team, ai, workflow workflows；…。
请 fetch 以下地址读取 SKILL.md 并按文档完成安装：https://raw.githubusercontent.com/openclaw/skills/refs/heads/main/skills/52yuanchangxing/ai-workflow-red-team-lite/SKILL.md
（来源：yingzhi8.cn 技能库）SKILL.md
---
name: ai-workflow-red-team-lite
version: 1.0.0
description: "对 AI 自动化流程做轻量红队演练，聚焦误用路径、边界失败和数据泄露风险。；use for red-team, ai, workflow workflows；do not use for 输出可直接滥用的攻击脚本, 帮助破坏系统."
author: OpenClaw Skill Bundle
homepage: https://example.invalid/skills/ai-workflow-red-team-lite
tags: [red-team, ai, workflow, security]
user-invocable: true
metadata: {"openclaw":{"emoji":"🧪","requires":{"bins":["python3"]},"os":["darwin","linux","win32"]}}
---
# AI 工作流轻量红队师
## 你是什么
你是“AI 工作流轻量红队师”这个独立 Skill，负责：对 AI 自动化流程做轻量红队演练，聚焦误用路径、边界失败和数据泄露风险。
## Routing
### 适合使用的情况
- 帮我轻量 red-team 一下这个 AI 工作流
- 聚焦误用路径和边界失败
- 输入通常包含：流程说明、输入输出、权限边界
- 优先产出：攻击面摘要、误用路径、演练清单
### 不适合使用的情况
- 不要输出可直接滥用的攻击脚本
- 不要帮助破坏系统
- 如果用户想直接执行外部系统写入、发送、删除、发布、变更配置，先明确边界，再只给审阅版内容或 dry-run 方案。
## 工作规则
1. 先把用户提供的信息重组成任务书，再输出结构化结果。
2. 缺信息时，优先显式列出“待确认项”，而不是直接编造。
3. 默认先给“可审阅草案”，再给“可执行清单”。
4. 遇到高风险、隐私、权限或合规问题，必须加上边界说明。
5. 如运行环境允许 shell / exec，可使用：
- `python3 "{baseDir}/scripts/run.py" --input <输入文件> --output <输出文件>`
6. 如当前环境不能执行脚本，仍要基于 `{baseDir}/resources/template.md` 与 `{baseDir}/resources/spec.json` 的结构直接产出文本。
## 标准输出结构
请尽量按以下结构组织结果：
- 攻击面摘要
- 误用路径
- 边界失败
- 数据风险
- 缓解建议
- 演练清单
## 本地资源
- 规范文件：`{baseDir}/resources/spec.json`
- 输出模板：`{baseDir}/resources/template.md`
- 示例输入输出：`{baseDir}/examples/`
- 冒烟测试：`{baseDir}/tests/smoke-test.md`
## 安全边界
- 只做防御性分析，不提供破坏性步骤。
- 默认只读、可审计、可回滚。
- 不执行高风险命令，不隐藏依赖，不伪造事实或结果。