Back to skill

Security audit

Cosdesign Standalone

Security checks for vulnerabilities and agentic risk

Overview

The skill is a disclosed design-analysis prompt generator, but it under-scopes arbitrary URL fetching and third-party Jina fallback risks.

Use this only with public, non-sensitive URLs. Do not analyze intranet, staging, authenticated, signed, or tokenized links unless you first remove the Jina fallback and add URL validation and prompt-injection guardrails.

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • 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
Findings (2)

T09 · Insecure Skill Coding Practices

Warning
Location
scripts/analyze.js:121
Finding

Unrestricted Network Target Delegation and Third-Party URL Disclosure

Content
View full analysis
` Site ${i+1}: ${u} WebFetch prompt: "Extract all visual design properties: colors, fonts, spacing, layout, border-radius, shadows from this page."`).join('\n')} ``` `scripts/define-style.js:39-40`: ```js 使用 WebFetch 获取 ${url},提取完整的视觉设计信息。 补充使用 WebSearch 搜索该品牌的设计语言相关信息(如有公开 Brand Guide)。 ``` `scripts/export.js:185-189`: ```js 使用 WebFetch 获取 ${url} 提取所有 CSS 属性:颜色、字体、间距、阴影、圆角、布局。 如果 WebFetch 返回内容不足,补充使用: WebFetch URL: https://r.jina.ai/${url} ``` ### Technical Analysis The scripts accept command-line values as URLs and insert them into prompts that direct a network-capable agent to invoke WebFetch. They do not validate: - The URL protocol. - Embedded usernames or passwords. - Loopback, private, link-local, or cloud metadata addresses. - Redirect destinations. - Hostnames resolving to private addresses. - Sensitive query parameters. - Whether the user has consented to sending the URL to Jina. Retrieving a public HTTP or HTTPS page is necessary for the declared design-analysis functionality. Allowing arbitrary network targets is not necessary. The Jina fallback creates an additional trust boundary because the complete user-supplied URL is placed after `https://r.jina.ai/`. A URL contain ...[truncated 2005 chars]
Remediation
View remediation

T01 · Skill Instruction Hijacking

Warning
Location
scripts/analyze.js:112
Finding

Prompt Injection Through Raw Argument Interpolation and Untrusted Web Content

Content
View full analysis
Remediation
View remediation
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (13)

Tp4

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

The declared description presents a functioning design-replication/analysis tool that visits websites and extracts concrete design-system information. The supplied code does not implement that behavior. Instead, it generates a textual prompt that tells another agent or tool how to do the analysis, including fallback instructions for WebFetch and Jina Reader. There is no network access, parsing of HTML/CSS, extraction logic, token generation, or comparison logic in the code itself. This is a material purpose/behavior mismatch: the description claims an operational analyzer, while the code is a prompt-template generator for such an analyzer.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

The declared description presents a functional tool that directly browses websites and produces concrete design-system outputs. However, the supplied code is explicitly a 'PROMPT GENERATOR ONLY' script. Its sole behavior is parsing command-line URLs and printing a comparison prompt template for an external agent to execute with WebFetch. That is a materially different primary purpose from actually crawling pages, extracting styles, and generating actionable design specs. While the printed prompt is related to design comparison, the implementation does not itself provide the core advertised capabilities.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

The declared description presents this skill as an operational design-replication analyzer that browses websites and extracts concrete design system information automatically. The supplied code does not implement that behavior. Instead, it is explicitly a 'PROMPT GENERATOR ONLY' script that formats and prints instructions for a downstream agent to do the fetching and analysis. Its primary purpose is therefore materially different: prompt generation rather than website analysis or design-token extraction. While the prompt text mentions WebFetch and WebSearch, the code itself never invokes those tools or accesses network resources. This is a clear description-behavior mismatch.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The description presents the skill as an operational website design-analysis and specification-generation tool. However, the supplied code is explicitly labeled 'PROMPT GENERATOR ONLY' and merely constructs a textual prompt instructing an external agent to use WebFetch or Jina and output one of several formats. There is no implemented fetching, parsing, extraction, transformation, or reporting logic beyond string templating and CLI argument handling. This is a material purpose mismatch: the declared skill claims direct automated analysis and export, while the code only scaffolds instructions for another system to do that work.

Content

No source excerpt is available for this finding.

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 181)May include surrounding context.

md
node scripts/compare.js <url1> <url2> [url3]

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The skill explicitly states it persists analysis history containing URLs, timestamps, and summaries, but does not warn users that these records may contain sensitive internal endpoints, project names, or confidential research targets. This creates a privacy and data-retention risk, especially in enterprise environments where browsing targets can themselves be sensitive.

Content

No source excerpt is available for this finding.

Sudo/Root Execution

Medium
Category
Privilege Escalation
Confidence
70% confidence
Finding

Commands invoke sudo or root privileges. Verify this elevated access is necessary and justified.

Content

Scanner excerpt · SKILL.md (reported line 131)May include surrounding context.

md
- 不写入用户项目目录(除非用户明确指定输出路径)

**权限级别:**
- 以当前用户身份运行,不需要 sudo
- 仅读取网页公开内容,不做任何写入或修改

**卸载:** 删除 skill 目录即可,无残留配置。

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The output template and instructions are written entirely in Chinese, including headings and formatting requirements, which steers the skill to produce Chinese-language results regardless of user preference. There is no opt-in, language selection flag, or justification that this skill is intentionally region-specific.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The script embeds the user-supplied URL into instructions that direct an agent to fetch the target through WebFetch and, on failure, through Jina Reader (https://r.jina.ai/...). This can disclose private or sensitive target URLs to third-party services without any warning, consent, or allowlist checks, which is especially risky if users analyze intranet, pre-release, or tokenized links.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The generated prompt is entirely in Chinese, including headings and task instructions, while the script accepts generic URLs and provides no option to choose language or locale. This imposes a specific language on downstream use without user opt-in, which matches the language/locale policy-violation category.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The generated prompt and all user-facing output are written in Chinese, including the title, instructions, and examples. This imposes a specific language/locale on downstream use without offering a choice or explaining a region-specific requirement, which matches the language-policy violation criteria.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

Multiple prompt templates and the main generated instructions are written to require Chinese output, such as '输出格式' and other Chinese-only directives. This imposes a specific language/locale behavior without user opt-in or a documented reason that the skill is region-specific.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The script generates instructions to fetch a user-supplied URL and, if that is insufficient, to send the same target to the r.jina.ai proxy. That can disclose sensitive URLs or internal endpoints to third parties without any explicit warning or consent, which is especially risky for a tool that may be used on private staging, intranet, or authenticated resources.

Content

No source excerpt is available for this finding.