Back to skill

Security audit

一键生成知乎高质量回答网页(可自由剪裁)

Security checks for vulnerabilities and agentic risk

Overview

ClawScan marks this skill suspicious with medium confidence because scanners report broad permission changes and file access beyond its stated research and webpage-generation purpose; review coverage is partial.

Consider avoiding installation until the permission-expansion instructions and out-of-project file access are reviewed. If you proceed, review authorization changes carefully and consider whether your research topics may contain information you do not want sent to external search services. Scanner coverage is marked partial, so the results do not establish that all referenced or runtime behavior was assessed.

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
Findings (19)

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
88% confidence
Finding

Downloading and installing the skill can be a supporting implementation detail, and the repository download is consistent with network access. However, copying files into ~/.claude/skills modifies resources outside the explicitly declared project-file write scope. The supplied chunk contains no research or webpage-generation logic; the behavior of the downloaded repository cannot be assessed from this chunk.

Content

No source excerpt is available for this finding.

Frames self-granting tool permissions as a routine fix for subagent permission denials.

High
Category
Prompt Injection
Confidence
98% confidence
Finding

Permission denials are treated as a reason for the agent to grant itself persistent access, rather than as an authorization boundary requiring review. This is particularly risky when general-purpose subagents consume external research content and can edit scripts subsequently executed by the workflow.

Content

No source excerpt is available for this finding.

HTML asset references can read files outside the permitted project and embed their contents in the distributable HTML.

High
Category
Excessive Agency
Confidence
99% confidence
Finding

Asset resolution accepts parent traversal, absolute paths, file:// references, and symlinks without verifying containment in the project images directory. An attacker-influenced HTML reference can therefore cause any file readable by the process to be base64-embedded in the deliverable; superficial structural checks do not reliably prevent sensitive-file inclusion.

Content

No source excerpt is available for this finding.

The automatic web-search workflow lacks a privacy warning about transmitting topic-derived queries to external search services.

Medium
Category
Output Handling
Confidence
88% confidence
Finding

The skill describes automatically performing extensive online searches based on the user's topic. Although the search activity and its scale are disclosed, the README does not warn that topic details included in search queries will leave the user's environment, which matters when a topic contains confidential or personal information.

Content

No source excerpt is available for this finding.

MCP server referenced without pinned version: 'npx skills'.

Medium
Category
MCP Rug Pull
Confidence
90% confidence
Finding

Although this is an installation CLI rather than an MCP server, the command uses npx to execute an unpinned npm package. A compromised future release could execute attacker-controlled code with the installing user's privileges; the unpinned skill repository also makes installed content mutable.

Content

No source excerpt is available for this finding.

MCP server referenced without pinned version: 'npx skills'.

Medium
Category
MCP Rug Pull
Confidence
90% confidence
Finding

This repeated installation command executes an unpinned npm CLI through npx, allowing the code executed during installation to change without changes to these instructions. The MCP-specific classification is inaccurate, but the underlying supply-chain exposure remains valid.

Content

No source excerpt is available for this finding.

The trigger list includes everyday phrases that can activate the skill outside its intended webpage-generation context.

Medium
Category
Trigger Abuse
Confidence
99% confidence
Finding

Generic triggers such as '怎么做' and '入门指南' extend beyond requests to generate research webpages. If used for automatic skill selection, they can initiate an unexpectedly expensive network-search and file-generation workflow for ordinary questions; the documented default of hundreds of searches amplifies this risk.

Content

No source excerpt is available for this finding.

Skill declares no tool scope ('permissions' or 'allowed-tools') but code capabilities were detected: file_read, file_write.

Medium
Category
MCP Least Privilege
Confidence
70% confidence
Finding

Without declared permissions the skill's intent is opaque and cannot be validated.

Content

No source excerpt is available for this finding.

The workflow instructs the agent to persistently expand its own tool permissions, although the skill's purpose is research and HTML generation.

Medium
Category
Excessive Agency
Confidence
97% confidence
Finding

The workflow directs the agent to create or modify persistent tool-authorization settings rather than obtain explicit approval for additional capabilities. Interpreter and curl wildcard rules, along with permission to edit an executable assembly script, can authorize more than narrowly scoped HTML production and amplify the effects of compromised subagents.

Content

No source excerpt is available for this finding.

Persistent agent permission changes are required without warning about their security impact.

Medium
Category
Output Handling
Confidence
97% confidence
Finding

Persistent permission expansion is presented as required setup without explaining that it changes the project's future authorization behavior. Users may consequently retain broader execution and write permissions after the research task ends.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
50% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · SKILL.md (reported line 627)May include surrounding context.

md
| 优先级 | 来源 | 判定标准 |
|--------|------|----------|
| 1 级 | 真实人物公开头像 | 公开可独立验证的直链(GitHub CDN `avatars.githubusercontent.com/u/<id>`、豆瓣影人页、雪球/掘金/丁香园等平台公开头像、个人官网) |
| 2 级 | DiceBear 风格化 SVG | `https://api.dicebear.com/7.x/{bottts-neutral|avataaars|notionists}/svg?seed=<英文短语>&backgroundColor=<hex>&radius=50`,三种风格交替配合不同配色 |
| 3 级 | 单字符占位 | `<div class="zh-avatar" aria-hidden="true">{首字}</div>`(仅在前两级均不可用时使用) |

**铁律**:

External Transmission

Medium
Category
Data Exfiltration
Confidence
50% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · SKILL.md (reported line 646)May include surrounding context.

md
curl -sI --max-time 5 "https://avatars.githubusercontent.com/u/1" | head -3

# 测 DiceBear API
curl -sI --max-time 5 "https://api.dicebear.com/7.x/bottts-neutral/svg?seed=test" | head -3

# 如有企业代理,补充 -x <代理地址>
curl -x <redacted-url> -sI --max-time 5 "https://avatars.githubusercontent.com/u/1" | head -3

External Transmission

Medium
Category
Data Exfiltration
Confidence
50% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · SKILL.md (reported line 682)May include surrounding context.

DiceBear SVG:直接保存为 .svg(远程 API 可能失效或限流)

curl -sL -o "./images/ch-02.svg"
"https://api.dicebear.com/7.x/avataaars/svg?seed=arduino-fan&backgroundColor=f4b400&radius=50"

text

**缓存规则**:

Restoring the skeleton can overwrite the existing page without a data-loss warning.

Medium
Category
Output Handling
Confidence
90% confidence
Finding

This recovery procedure applies when index.html still exists but rerunning assembly has no effect. Restoring the skeleton can discard edits in that file, yet the instructions neither warn about losing those edits nor require a backup before replacement.

Content

No source excerpt is available for this finding.

The template requires at least 11,000 Chinese characters without offering a language choice or user opt-in.

Medium
Category
Output Handling
Confidence
94% confidence
Finding

The requirement for “有效中文字数” and mandatory expansion until the threshold is met impose Chinese-language content, rather than merely presenting instructions in Chinese. The template provides neither a language-selection option nor a documented justification for restricting the output language.

Content

No source excerpt is available for this finding.

The filename rule requires Chinese topics to be translated into English without user opt-in.

Low
Category
Output Handling
Confidence
88% confidence
Finding

The naming rule explicitly mandates translating Chinese topic names into English for the generated HTML filename. Although it specifies an ASCII filename format, it offers no language choice or explicit justification for requiring English rather than another ASCII-compatible naming scheme.

Content

No source excerpt is available for this finding.

The output requirement mandates Chinese-language text without offering a language choice.

Low
Category
Output Handling
Confidence
92% confidence
Finding

The requirement for at least 100,000 effective Chinese characters makes Chinese output a condition of the default artifact, rather than merely using Chinese to document the skill. No user language choice is offered, and reproducing Zhihu's visual style does not itself require Chinese-language content.

Content

No source excerpt is available for this finding.

The filename convention mandates English translation without a language opt-in.

Low
Category
Output Handling
Confidence
88% confidence
Finding

The naming rule explicitly requires translating the topic into English and rejects pinyin, irrespective of the user's selected writing language. An ASCII filename constraint does not itself justify requiring a particular natural language.

Content

No source excerpt is available for this finding.

Documentation promises interactive HTML repair, but --fix exits without scanning or writing repaired files.

Low
Category
Excessive Agency
Confidence
99% confidence
Finding

L011 documents --fix as an interactive repair mode that produces draft*_fixed.html files. In contrast, L132–L134 only print that repair is unimplemented and return an error before scanning, directly contradicting the documented behavior.

Content

No source excerpt is available for this finding.