Ai Fact Checker

🔍 AI 事实核查 - 自动验证大模型输出准确性，检测 hallucination 幻觉编造内容，联网搜索交叉验证给出可信度评分和修正。解决大模型一本正经胡说八道的痛点。

Security Scan

The skill's purpose and files are mostly coherent for a fact-checker, but the runtime instructions/code include an unsanitized shell invocation and SKILL.md contains suspicious unicode control characters—these are security-relevant issues that should be fixed or reviewed before installing.

Purpose & Capability

Name/description, SKILL.md, and code all describe the same behavior: extract factual statements and use OpenClaw's web_search to verify them. The skill requests no secrets, no unrelated binaries, and the declared dependency on the built-in web_search tool matches the implementation (it calls an openclaw web_search CLI).

Instruction Scope

The SKILL.md and scripts instruct the agent to run a CLI web_search and parse results — that's expected — but scripts/fact-check.js uses child_process.execSync to run: openclaw tool web_search --query "<user text>". The user-supplied statements are interpolated into a shell command with only double-quote escaping; other shell metacharacters (`, $, ;, &, |, $(), etc.) are not escaped, creating a command-injection risk. Additionally, the SKILL.md contains unicode control characters (scanner flagged 'unicode-control-chars'), which can be used for prompt-injection or to confuse parsers; that is unexpected for a benign fact-checker and merits manual review.

Install Mechanism

No install spec is provided (instruction-only), which is low-risk for supply-chain downloads. However, the skill bundle does include JavaScript files that will be present once installed; since these are executed locally when you run tests or the script, you should review them. There is no remote download URL in the install, which reduces supply-chain concerns.

Credentials

The skill requests no environment variables or credentials, consistent with a public web-search based fact-checker. The code does not attempt to read environment secrets. The only external dependency is the OpenClaw CLI/web_search tool; ensure that tool's behavior and returned JSON format are trusted/expected.

Persistence & Privilege

The skill does not request always:true, does not claim to modify other skills or global agent settings, and is user-invocable by default. It does not attempt to persist credentials or alter system-wide configuration.

Scan Findings in Context

[unicode-control-chars] unexpected: SKILL.md was flagged for unicode control characters. These are not needed for a fact-checker's instructions and can be used to manipulate parsers or hide adversarial content; manual inspection and removal are recommended.

What to consider before installing

What to consider before installing: - Don't install blindly. The skill's purpose is coherent, but the code contains a command-injection vulnerability: scripts/fact-check.js builds a shell command by interpolating user text into execSync(openclaw tool web_search --query "..."). An attacker-controlled input could inject shell metacharacters (e.g., `;`, `&&`, `$(...)`, backticks) and execute arbitrary commands on your host when the script runs. - If you must use it: run it only in a sandboxed or isolated environment (container/VM) until you fix the code. - Fixes to request or apply before running: - Replace execSync with a safer invocation that passes arguments as an argv array (avoid the shell) or use the OpenClaw SDK/API instead of shelling out. For Node, use child_process.spawn or execFile with an array of arguments to avoid shell interpolation. - Properly validate/escape user input. At minimum, reject inputs containing suspicious shell metacharacters or sanitize them using a robust escaping library. - Add robust error handling for the web_search output before JSON.parse to avoid crashes or injection via malformed output. - Remove any unicode control characters from SKILL.md and verify there are no hidden/invisible characters intended to manipulate downstream parsers. - Verify the openclaw CLI behavior: confirm that openclaw tool web_search returns well-formed JSON and that invoking it from scripts is allowed and safe in your environment. - If you don't want to run code from this unknown source, consider implementing the same logic yourself or using a vetted fact-checking skill/utility. Overall: the skill is functionally coherent but contains security issues (shell injection risk + suspicious control characters). Treat it as suspicious until the above issues are resolved.

🔍 AI Fact Checker - AI 事实核查

自动验证 AI 输出信息的真实性，检测幻觉（hallucination）编造内容，联网搜索交叉验证，给出可信度评分，并自动修正错误信息。

✨ 痛点解决

大模型经常一本正经地胡说八道：编造不存在的论文、数据、人名、事件、网址。这个技能帮你：

1. 自动提取 - 智能识别文本中的事实性陈述，排除观点和主观判断
2. 交叉验证 - 自动联网搜索每个陈述，对比权威来源
3. 可信度评分 - 给出 ✅ 可信 / ⚠️ 部分存疑 / ❌ 大概率错误
4. 自动修正 - 如果发现错误，基于搜索结果给出正确信息
5. 引用来源 - 附上搜索结果链接，方便你进一步核实

🚀 使用方法

基本核查

请用 ai-fact-checker 验证这段信息：
OpenClaw 发布于 2025 年，ClawHub 上有超过 10000 个社区技能。

对话中快速核查

当你对 AI 回答某句话存疑时：

fact-check 这句话是否正确：OpenClaw 的作者是 Peter Steinberger

批量核查整个文档

fact-check-document ./path/to/your-document.md

⚙️ 工作流程

原始文本 → 提取事实陈述 → 逐个联网搜索 → 关键词匹配评分 → 生成核查报告

1. 智能提取 - 过滤掉观点，只保留可验证的事实陈述
2. 独立搜索 - 每个陈述单独搜索，保证准确性
3. 匹配算法 - 根据关键词覆盖率计算可信度分数
4. 结构化报告 - 清晰展示每个陈述的验证结果

📊 评分标准

评分范围	标识	说明
90-100	✅ 可信	信息与多个权威来源一致
60-89	⚠️ 部分可信	核心信息得到验证，细节无法确认
30-59	⚠️ 存疑	找不到足够验证来源或信息部分冲突
0-29	❌ 错误	信息与权威来源矛盾，确认为编造

💡 示例输出

# 🧐 AI 事实核查报告

**原始文本:**
OpenClaw 发布于 2025 年，现在 ClawHub 上已经有超过 10000 个社区技能。作者是 Peter Steinberger。

---

## 1. 核查: "OpenClaw 发布于 2025 年，现在 ClawHub 上已经有超过 10000 个社区技能。作者是 Peter Steinberger。"

- **评分:** 45/100
- **结论:** ⚠️ 存疑
- **原因:** 部分信息与搜索结果不符

**参考来源:**
1. [OpenClaw GitHub - README](https://github.com/openclaw/openclaw)
2. [ClawHub - Official Skill Registry](https://clawhub.ai/skills)

---

**整体评分:** 45/100 - ⚠️ 整体存疑，建议核实

**修正:**
- OpenClaw 发布于 **2024 年**（不是 2025）
- ClawHub 上目前收录 **~5700+** 个技能（不是 10000+）
- ✅ 作者确实是 **Peter Steinberger**

🔧 依赖

• 需要 OpenClaw 内置 web_search 技能（默认自带，不需要额外安装）
• 不需要额外 API 密钥，使用现有搜索配额

📦 安装

npx clawhub install ai-fact-checker

🧪 测试

cd ai-fact-checker
node test-run.js

👨‍💻 Author

rudagebil11-jpg

📄 License

MIT