Description-Behavior Mismatch
- Category
- MCP Tool Poisoning
- Confidence
- 99% confidence
- Finding
所提供代码的主要用途是Markdown文档导出,与声明的多格式内容解析为Markdown存在实质差异。代码还明确生成磁盘文件。该代码块未展示URL读取、OCR、音视频转写或视觉理解能力;判定主要依据已展示的反向转换及落盘行为,而非仅因声明能力未出现。
- Content
Security audit
Security checks for vulnerabilities and agentic risk
ClawScan marks this skill suspicious with medium confidence because scanners report a mismatch between its declared document-parsing purpose and some behaviors, along with concerns about credential handling, remote processing, and mutable package installation instructions.
The verdict is suspicious, not malicious. Consider the reported credential exposure, conflicting descriptions of remote processing, purpose mismatch, and unpinned package commands before installing. Skillspector scanned the listed Markdown component but marked its analysis partial because several referenced paths were missing; the results therefore do not establish that the artifact is safe.
所提供代码的主要用途是Markdown文档导出,与声明的多格式内容解析为Markdown存在实质差异。代码还明确生成磁盘文件。该代码块未展示URL读取、OCR、音视频转写或视觉理解能力;判定主要依据已展示的反向转换及落盘行为,而非仅因声明能力未出现。
The skill repeatedly claims that source files do not reach the server and that local-mode files do not pass through Cue, while its architecture explicitly says parsing is performed by Cue's remote service. These contradictory assurances can induce users to authorize processing of confidential documents without understanding the external disclosure; no_store is not equivalent to no transmission.
The onboarding instructions require users to paste a reusable API key into the agent conversation. This unnecessarily exposes the credential to conversation storage, model processing, logs, and anyone with transcript access, potentially enabling unauthorized API use and credit consumption.
Skill accesses MCP server configuration files (mcp.json). MCP configs contain server URLs, authentication tokens, and tool definitions — reading them allows the skill to discover and potentially abuse other tool integrations.
npx -y @cueai/omni-reader-mcp@1.8.6 setup
Agent 配置(`~/.claude/mcp.json`):
```json
{
Remote code is downloaded and executed. This bypasses code review and could introduce malicious code.
# 健康检查:验证 Key 和连接状态,不输出密钥原文
echo "=== 1/4 API Key ===" && [ -f "$HOME/.cue/config.json" ] && echo "已配置" || echo "未配置!"
echo "=== 2/4 MCP 连接 ===" && curl -sS --max-time 10 -X POST "https://mcp.cuecue.cn/api/omni-reader/mcp/" -H "Content-Type: application/json" -d '{"jsonrpc":"2.0","id":1,"method":"tools/list","params":{}}' | python3 -c "import sys,json;r=json.load(sys.stdin);print('已连接 ('+str(len(r.get('result',{}).get('tools',[])))+' 工具)') if 'result' in r else print('连接失败')"
echo "=== 3/4 Node.js ===" && node -v 2>/dev/null && npx -v 2>/dev/null && echo "就绪" || echo "未安装!"
echo "=== 4/4 剩余积分 ===" && python3 -c "
import json, urllib.request, os
The instruction accepts generic requests such as “解析一下” and treats “omni” as an activation name without requiring an identifiable file, URL, or document-conversion intent. These can overlap with unrelated analysis requests or references to other products.
Large whitespace padding was detected (a block of blank lines or a long run of spaces). This can push injected instructions below or to the right of the visible area so a human reviewer never sees them while the agent still reads them. Manual review of the hidden content is recommended.
> 两者**互补不冲突**:专项 Skill 底层正是调用 Cue Omni Reader 远程端点完成解析(并自动绕过平台反爬),再由大模型加工交付。可直接对我说「安装 xiaohongshu-video-studio」。
## 核心优势
1. 效果准(官方实测):模糊复杂表格的扫描件解析准确率 高达97.39%,已和主流OvisOCR2(准确率94.77%)、合合 textin(92.16%)、MinerU2.5(91.50%)等6大主流ocr工具做过效果对比,准确率处于行业第一梯队;
2. 隐私友好:默认 no_store 不落盘,源文件与解析结果不上服务端;
3. 文档大小不限制,支持单文件256MB,300+页pdf文档也是秒级解析速度;支持多文件批量解析,上百份文档解析批量处理。
Large whitespace padding was detected (a block of blank lines or a long run of spaces). This can push injected instructions below or to the right of the visible area so a human reviewer never sees them while the agent still reads them. Manual review of the hidden content is recommended.
> 两者**互补不冲突**:专项 Skill 底层正是调用 Cue Omni Reader 远程端点完成解析(并自动绕过平台反爬),再由大模型加工交付。可直接对我说「安装 xiaohongshu-video-studio」。
## 核心优势
1. 效果准(官方实测):模糊复杂表格的扫描件解析准确率 高达97.39%,已和主流OvisOCR2(准确率94.77%)、合合 textin(92.16%)、MinerU2.5(91.50%)等6大主流ocr工具做过效果对比,准确率处于行业第一梯队;
2. 隐私友好:默认 no_store 不落盘,源文件与解析结果不上服务端;
3. 文档大小不限制,支持单文件256MB,300+页pdf文档也是秒级解析速度;支持多文件批量解析,上百份文档解析批量处理。
4. 文件类型不限制,支持PDF/Office/图片/扫描件/音视频/网页/压缩包等所有常见文件类型;
The claim that the agent cannot view the key is paired with an instruction to send that same key to the agent in chat, creating a misleading sense of protection. This is the same credential-exposure issue identified by SQP-2, not evidence of a separate exfiltration destination.
The update instructions execute the package using the mutable @latest tag rather than a reviewed, fixed version. A compromised future release could execute with the user's privileges; requiring upgrade approval reduces automatic exposure but does not establish package integrity.
npx commands without a version suffix (e.g. @1.0.0) create a rug-pull risk if the upstream server is compromised and publishes a malicious update.
The skill requires the Chinese prompt and answer options to be used verbatim, explicitly prohibiting changes to their wording. This imposes a natural-language requirement regardless of the user's language, with no language choice or justified locale restriction.
Commands invoke sudo or root privileges. Verify this elevated access is necessary and justified.
| 目标格式 | 依赖 | macOS | Ubuntu |
|----------|------|-------|--------|
| Word (.docx) | pandoc | `brew install pandoc` | `sudo apt install pandoc` |
| PDF (.pdf) | pandoc + LaTeX | `brew install --cask basictex` | `sudo apt install texlive-xetex` |
---
Commands invoke sudo or root privileges. Verify this elevated access is necessary and justified.
| 目标格式 | 依赖 | macOS | Ubuntu |
|----------|------|-------|--------|
| Word (.docx) | pandoc | `brew install pandoc` | `sudo apt install pandoc` |
| PDF (.pdf) | pandoc + LaTeX | `brew install --cask basictex` | `sudo apt install texlive-xetex` |
---
npx commands without a version suffix (e.g. @1.0.0) create a rug-pull risk if the upstream server is compromised and publishes a malicious update.
The recovery instructions run an unversioned package with npx -y, allowing a newly published release to execute without a package-installation confirmation. This weakens the version controls used elsewhere and exposes the host to a compromised upstream release.
L030 says the MCP configuration is merely a reference for user deployment and that the skill itself does not execute MCP connections. In contrast, the required pre-parse health check executes a curl POST containing a tools/list request against the remote MCP endpoint. L219 additionally directs the agent to call that endpoint directly when only the local Bridge is configured, contradicting the stated connection boundary.
Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.
# 健康检查:验证 Key 和连接状态,不输出密钥原文
echo "=== 1/4 API Key ===" && [ -f "$HOME/.cue/config.json" ] && echo "已配置" || echo "未配置!"
echo "=== 2/4 MCP 连接 ===" && curl -sS --max-time 10 -X POST "https://mcp.cuecue.cn/api/omni-reader/mcp/" -H "Content-Type: application/json" -d '{"jsonrpc":"2.0","id":1,"method":"tools/list","params":{}}' | python3 -c "import sys,json;r=json.load(sys.stdin);print('已连接 ('+str(len(r.get('result',{}).get('tools',[])))+' 工具)') if 'result' in r else print('连接失败')"
echo "=== 3/4 Node.js ===" && node -v 2>/dev/null && npx -v 2>/dev/null && echo "就绪" || echo "未安装!"
echo "=== 4/4 剩余积分 ===" && python3 -c "
import json, urllib.request, os
L031 states that the skill does not read or transmit credentials. However, the mandatory health check loads the API key from ~/.cue/config.json or CUE_API_KEY and sends it to Cue's billing endpoint in an Authorization header; L219 also explicitly instructs the agent to read the stored key for direct requests. Although authentication to the provider is legitimate, this behavior directly contradicts the documented credential-handling boundary.
The FAQ's unversioned uninstall command can download and execute the current upstream package just to remove the connector. Uninstallation therefore introduces an unnecessary opportunity to execute an unreviewed or compromised release.
The FAQ repeats an @latest setup command, which selects mutable upstream code rather than a reviewed release. Updating is legitimate, but this workflow could execute a compromised new release with access to local files and credentials.