Skill flagged — suspicious patterns detected

ClawHub Security flagged this skill as suspicious. Review the scan results before using.

Find Everything

v0.1.0

跨平台资源搜索编排器。搜索 skill、MCP 服务器、提示词模板、开源项目。 覆盖 skills.sh、ClawHub、SkillHub、AI Skills Show、MCPServers.org、 prompts.chat、GitHub 等 14+ 个聚合站。 触发场景:用户说"找个xxx工具"、"有没有xx...

0· 277·0 current·0 all-time

Install

OpenClaw Prompt Flow

Install with OpenClaw

Best for remote or guided setup. Copy the exact prompt, then paste it into OpenClaw for zadanthony/find-everything.

Previewing Install & Setup.
Prompt PreviewInstall & Setup
Install the skill "Find Everything" (zadanthony/find-everything) from ClawHub.
Skill page: https://clawhub.ai/zadanthony/find-everything
Keep the work scoped to this skill only.
After install, inspect the skill metadata and help me finish setup.
Use only the metadata you can verify from ClawHub; do not invent missing requirements.
Ask before making any broader environment changes.

Command Line

CLI Commands

Use the direct CLI path if you want to install manually and keep every step visible.

OpenClaw CLI

Bare skill slug

openclaw skills install find-everything

ClawHub CLI

Package manager switcher

npx clawhub@latest install find-everything
Security Scan
VirusTotalVirusTotal
Benign
View report →
OpenClawOpenClaw
Benign
high confidence
Purpose & Capability
Name/description describe a search/aggregation tool; SKILL.md, registry.json, and references/ files implement that behavior. Requested capabilities (reading registry.json, running CLIs like gh/npx if available, calling MCP tools, doing websearch/WebFetch) are coherent with a cross-source search aggregator.
Instruction Scope
Instructions include running CLI commands (npx, gh, clawhub) and MCP/tool calls, performing WebFetch of remote repo files, installing packages into a temporary directory for scanning, and optionally modifying registry.json when user confirms. This is expected for a search/installer workflow, but it implies executing arbitrary third‑party code during explicit 'install' flows — user confirmation and review of scan output are relied upon.
Install Mechanism
There is no install spec (instruction-only), and the only included code is a deterministic scanner (scripts/security_scan.py) plus registry and checklist references. No downloads/install hooks are executed automatically by the skill itself; remote installs happen only when the user asks to 'install' a discovered resource (via npx/gh, etc.).
Credentials
The skill declares no required environment variables, credentials, or config paths. The runtime instructions reference checking for local CLIs/tools and using MCP tools available in-session — which is proportionate to the stated functionality.
Persistence & Privilege
always:false and no global privileges requested. The skill does write to its own registry.json when the user approves adding new sources (modifies files within its bundle). It may initiate installs/webfetch on user request — these are expected but carry execution risk that the user should acknowledge.
Scan Findings in Context
[ignore_instructions] expected: Pre-scan flagged 'ignore-previous-instructions' pattern in SKILL.md. The SKILL.md and references/security-checklist explicitly discuss prompt-injection patterns (including that phrase) as part of detection guidance, so this pattern appears to be referenced for defensive scanning rather than an attempt to override the agent.
Assessment
This skill looks internally consistent with its stated purpose, but be aware: (1) it will run local CLIs (gh, npx, clawhub) and MCP tools if available and will perform webfetches of remote repositories; (2) if you ask it to 'install' or to run a deep scan, it may download/install third‑party packages into a temporary directory (which can execute untrusted code) — review the security_scan.py results and the fetched files before proceeding; (3) the skill can modify its own registry.json to add new sources with your confirmation; (4) the SKILL.md references prompt‑injection patterns (flagged by pre-scan) for defensive purposes, but you should still review the skill and consider disabling autonomous invocation or requiring manual confirmation for installs or any operation that runs external code or writes files.
!
references/security-checklist.md:11
Prompt-injection style instruction pattern detected.
About static analysis
These patterns were detected by automated regex scanning. They may be normal for skills that integrate with external APIs. Check the VirusTotal and OpenClaw results above for context-aware analysis.

Like a lobster shell, security has layers — review code before you run it.

latestvk97cy03kc6cte2zfbkcmzergdx8361g7
277downloads
0stars
1versions
Updated 22h ago
v0.1.0
MIT-0

find-everything:跨平台资源搜索

触发条件

  1. 显式调用/find-everything {query}
  2. 自动检测:用户问"有没有xxx工具"、"帮我找个xxx"、"有没有skill能..."
  3. 主动推荐:基于对话上下文判断用户在持续做某类任务且缺少相关工具(best-effort 启发式,同会话最多 1 次)

执行流程

Step 1: 意图分类

将用户查询分类为一个或多个类别:

  • skill: Agent Skills(可安装技能)
  • mcp: MCP Servers
  • prompt: 提示词模板/角色扮演/图片生成
  • repo: GitHub 开源项目

模糊需求同时搜索多个类别。从自然语言中提取核心搜索关键词。

Step 2: 读取注册表并路由

读取 references/registry.json。若读取失败,使用以下硬编码最小源:

  • skills-sh: npx skills find {query}
  • github: gh search repos {query} --sort stars --limit 10 --json name,owner,description,url,stargazersCount

筛选规则:

  1. 只选 enabled: true 的源
  2. 只选 category 包含目标类别的源
  3. 检查 requires 是否满足:
    • requires 为 CLI 名称(如 npxgh)→ 用 which 检查
    • requiresmcp: 开头(如 mcp:prompts-chat)→ 检查对应 MCP tool 是否在当前会话可用(尝试 ToolSearch 或直接调用)
  4. 不可用的源跳过,记录到 skipped_sources 列表,附带 install_hint(如有)

Step 3: Tier 1 搜索

同一条回复中发起多个独立 tool 调用实现并行:

cli 类型:用 Bash tool 执行 command 字段({query} 替换为实际关键词),15 秒超时。 mcp 类型:调用 registry.json 中 tool 字段指定的 MCP tool,参数取 tool_params({query} 替换为实际关键词)。例如:

  • search_prompts({ query: "blockchain analyst", limit: 10 }) → 返回 prompt 列表
  • search_skills({ query: "search aggregator", limit: 10 }) → 返回 skill 列表
  • 支持的额外过滤参数(按需使用):type(TEXT/STRUCTURED/IMAGE)、categorytag skill 类型:调用对应已安装 skill。

若并行不可用,按优先级执行:skills-sh > github > clawhub > 其他。

Step 4: 结果数量预判

基于标题/描述的关键词匹配做轻量判断(非完整 LLM 评估):

  • ≥3 条匹配度高 → 跳到 Step 6(末尾注明"搜索更多源可获取更多结果")
  • <3 条或匹配度低 → 继续 Step 5

Step 5: Tier 2/3 搜索

Tier 2:将同类别 Tier 2 源合并为 1 次 WebSearch:

{query} {category关键词} site:skillhub.club OR site:aiskillsshow.com OR site:skillsmp.com

若合并查询返回 <2 条,退回对 Top 2-3 优先源逐个 site: 搜索。

Tier 3(仍不足时):不带 site: 限定的广域 WebSearch。

Step 6: 结果评估

对每条结果判断相关性:

  • 高相关:保留,排在前面
  • 中相关:保留,排在后面,标注"可能相关"
  • 不相关:丢弃

对 Top 3-5 条高相关结果,可选 WebFetch 补充详情。

Step 7: 安全快筛

对所有保留结果做元数据安全标注:

  • [SAFE]:来自注册表已知平台 + 安装量 >100
  • [CAUTION]:安装量低、来源不明、或 Tier 3 发现
  • [RISK]:名称疑似 typosquat,或其他红旗信号

Step 8: 去重 + 排序

同一工具出现在多个源 → 合并为一条,标注所有来源。 合并时优先保留安装量/star 数最高的来源作为主展示(同一站点可能有 Tier 1 CLI 和 Tier 2 WebSearch 两条结果,合并时保留 Tier 1 数据)。 排序:相关度 > 安全等级 > 安装量/star 数。

Step 9: 展示推荐

格式:

找到 N 个相关结果(来自 X 个源)

1. [名称]  [SAFE]
   来源: skills.sh | 安装量: 1.2K
   简介: ...
   推荐理由: ...

2. [名称]  [CAUTION]
   来源: WebSearch (skillhub.club) | star: 50
   简介: ...
   注意: 来源非直接 API,建议安装前审查源码

---
未覆盖的搜索源: clawhub CLI(未安装)
提示: npm i -g clawhub

注意:同一依赖缺失提示每会话只展示一次,避免重复打扰。

Step 10: 用户后续操作

用户说**"看看"/"详细"** → WebFetch 详情页展示更多信息。

用户说**"安装"/"用这个"** → 触发深度安全扫描:

1. 获取资源内容(按类型):

资源类型扫描目标获取方式
Skill (skills.sh/clawhub)SKILL.md + scripts/ 目录优先 WebFetch GitHub 源码(从搜索结果中的 repo URL 获取);若不可访问,npx skills add <name> 安装到临时目录后读取
MCP Serverpackage.json + 入口文件WebFetch npm registry 页面或 GitHub README,提取入口文件路径后 WebFetch
GitHub RepoREADME.md + 主要脚本文件gh api repos/{owner}/{repo}/contents 获取文件列表,WebFetch 关键文件
Prompt提示词文本本身通常搜索结果中已包含完整文本,无需额外获取

限制: 单次扫描最多 50KB。超出只扫 SKILL.md + 入口文件。

2. 运行安全扫描(路径相对于本 SKILL.md 所在目录解析):python3 scripts/security_scan.py <file> [--check-name <name> --known-skills references/known_skills.txt] 3. 读取 references/security-checklist.md,结合 security_scan.py 结果做 LLM 上下文评估 4. 输出量化评分(0-100)和风险定级 5. [SAFE] → 执行安装;[CAUTION] → 展示详情让用户确认;[RISK] → 明确提示风险,不自动安装

Step 11: 新站点发现

Tier 3 搜索发现了不在 registry.json 中的优质资源站 → 提示用户: "发现新站点 xxx.com,内容相关度高。要加入搜索注册表吗?" 用户确认 → 在 registry.json 的 sources 数组中追加新条目。

错误处理

场景处理
CLI 超时(15s)跳过该源,继续其他
WebSearch 错误/限流跳过 Tier 2,展示 Tier 1 结果 + 提示稍后重试
security_scan.py 崩溃退回 LLM-only 评估,标记"自动扫描未完成"
所有 Tier 1 不可用直接进 Tier 2,注明"主要搜索源暂不可用"
所有源失败告知用户搜索失败,建议检查网络

主动推荐逻辑

触发条件(低频,同会话最多 1 次):

  • 对话上下文显示用户持续在做某类任务(如调试、前端开发、数据处理)
  • 且当前未见明显相关的 skill/MCP 在使用

推荐格式: "[发现] 你正在做 xxx,有一些工具可能有帮助,要搜索看看吗?"

Comments

Loading comments...