Back to skill

Security audit

AI无障碍合规守护者

Security checks for vulnerabilities and agentic risk

Overview

This skill is a coherent accessibility-audit workflow, but users should only run it on approved pages because it captures rendered page content for AI analysis.

Install only if you are comfortable with a workflow that opens supplied pages in Chromium, executes page JavaScript, captures screenshots and DOM including iframe content, and analyzes that content with external skill dependencies. Avoid private, authenticated, internal, regulated, or secret-bearing pages unless you have explicit approval and understand the data handling path.

Vulnerability Patterns
  • Insecure DependenciesIntroduces malicious components through unsafe dependency sources
  • 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
Findings (2)

other

Warning
Location
workflow.json:26
Finding

Untrusted Webpage DOM Is Passed Directly to an AI Analyzer

Content
View full analysis

Vulnerability Details

File Location: workflow.json:26-45
Vulnerability Type: Indirect prompt injection
Risk Level: Medium

Vulnerable Code

json
      "extract": "dom"
    },
    "output": {
      "dom": "${DOM_CONTENT}",
      "screenshot": "${PAGE_SCREENSHOT}",
      "title": "${PAGE_TITLE}",
      "links": "${PAGE_LINKS}"
    },
    "description": "使用真实 Chromium 浏览器完整渲染页面(含 JS 执行),截图并提取完整 DOM 结构。捕获 Shadow DOM、懒加载内容、iframe 内内容。"
  },
  {
    "order": 2,
    "name": "WCAG合规AI分析",
    "skill": "summarize",
    "input": {
      "source": "${DOM_CONTENT}",
      "mode": "wcag-analysis",
      "level": "${LEVEL:-AA}",
      "standard": "WCAG 2.1",

Technical Analysis

The workflow extracts the complete DOM from an attacker-controlled target page and supplies it directly to the summarize AI component as its analysis source. No explicit trust boundary, content sanitization, prompt-injection defense, or instruction/data separation is configured before this transfer.

Because the target website controls the DOM, it can embed natural-language instructions in visible elements, hidden elements, comments, Shadow DOM content, or iframe content. If the AI analyzer interprets these instructions as directives rather than as untrusted evidence, the page can influence the generated accessibility findings.

The available configuration does not establish that webpage content can invoke tools or execute commands. Therefore, the demonstrated risk concerns the integrity of the WCAG analysis and generated report rather than direct system compromise.

Attack Path

  1. An attacker creates or modifies a webpage containing prompt-like instructions designed to influence an AI analyzer.
  2. A user requests an accessibility audit of the malicious URL.
  3. agent-browser loads the page, executes its JavaScript, and extracts the complete DOM, including attacker-controlled content.
  4. The wor ...[truncated 952 chars]
Remediation
View remediation

Remediation Suggestions

  1. Treat all extracted DOM content as untrusted data and place it inside explicit, strongly delimited data sections.
  2. Add a higher-priority instruction stating that directives, requests, or policies found inside the target DOM must never be followed.
  3. Transform the DOM into a constrained accessibility-analysis representation rather than submitting unrestricted page text.
  4. Remove scripts, comments, metadata, hidden prompt-like content, and other elements that are unnecessary for the selected WCAG checks.
  5. Require the analyzer to return a schema-constrained result containing only expected finding fields, WCAG identifiers, evidence, severity, and remediation.
  6. Validate each reported issue against concrete DOM evidence before accepting it.
  7. Escape or sanitize all analyzer-generated text before passing it to card-renderer.
  8. Apply output length limits and reject unexpected instructions, markup, URLs, or fields in generated findings.
  9. Where practical, use deterministic accessibility tools for machine-verifiable checks and reserve the AI model for explanation and remediation guidance.

T08 · Insecure Dependencies

Note
Location
workflow.json:96
Finding

External Skill Dependencies Are Not Reproducibly Pinned

Content
View full analysis

Vulnerability Details

File Locations: SKILL.md:14-17; workflow.json:96-109
Vulnerability Type: Insecure dependency and supply-chain configuration
Risk Level: Low

Vulnerable Code

SKILL.md:14-17:

yaml
dependencies:
  - agent-browser
  - summarize
  - card-renderer

workflow.json:96-109:

json
  "dependencies": {
    "agent-browser": {
      "purpose": "页面完整浏览器渲染(含 JS 执行),截图 + 提取 DOM",
      "version": ">=1.0.0"
    },
    "summarize": {
      "purpose": "AI WCAG 合规分析,输出问题清单 + 修复建议",
      "version": ">=1.0.0"
    },
    "card-renderer": {
      "purpose": "生成可视化无障碍合规报告卡片",
      "version": ">=1.0.0"
    }
  },

Technical Analysis

The project delegates all operational behavior to three external Skills but does not identify trusted package sources, exact reviewed versions, cryptographic hashes, signatures, or a dependency lockfile. The version constraint >=1.0.0 permits any current or future compatible version to be selected.

Consequently, the effective implementation executed by the workflow can change after this project has been reviewed. A compromised dependency publisher, registry compromise, namespace substitution, or malicious future release could introduce behavior that is absent from the audited files.

The project files do not prove that any dependency is currently malicious. This finding concerns the absence of reproducible dependency resolution and integrity controls.

Attack Path

  1. An attacker compromises a dependency publisher or distribution source, substitutes a similarly named component, or publishes a malicious version accepted by >=1.0.0.
  2. A deployment or subsequent dependency resolution selects the attacker-controlled release.
  3. The workflow invokes the altered dependency under the trusted names agent-browser, summarize, or card-renderer.
  4. The malicious component receives the data and capabilities associa ...[truncated 885 chars]
Remediation
View remediation

Remediation Suggestions

  1. Replace broad >=1.0.0 constraints with exact, reviewed dependency versions.
  2. Record dependency resolution in a lockfile or equivalent immutable manifest.
  3. Declare the trusted registry, repository, publisher identity, and package namespace for every Skill.
  4. Verify cryptographic hashes or publisher signatures before loading dependencies.
  5. Reject unsigned, substituted, or unexpectedly updated components.
  6. Review dependency updates before changing pinned versions and repeat the security audit after each update.
  7. Apply least-privilege controls so each dependency receives only the data and runtime capabilities needed for its stage.
  8. Restrict browser network access where possible and prevent analyzer and renderer components from making unnecessary outbound requests.
  9. Maintain an allowlist of approved dependency identities and versions in deployment policy.
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
Findings (5)

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The README describes loading arbitrary third-party URLs in a real Chromium browser, executing JavaScript, and capturing screenshots and full DOM content, but it does not warn users about those behaviors. This can lead users to submit sensitive or internal URLs without understanding that page contents, scripts, rendered state, and embedded data may be fetched and exposed to downstream components for analysis and reporting.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The skill explicitly states that rendered pages, screenshots, DOM structure, and even iframe content are extracted and then analyzed by an external AI service, but it does not clearly warn users that potentially sensitive page content will be transmitted off-platform. This creates a real privacy and data-handling risk, especially for internal pages, authenticated sessions, or regulated environments where DOM contents and screenshots may contain secrets, personal data, or confidential business information.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The workflow uses a real browser to visit arbitrary user-supplied URLs and captures full DOM content, screenshots, titles, and links, which can include sensitive information such as authenticated page data, personal content, internal application state, or embedded secrets in rendered HTML. In this accessibility-audit context the collection is functionally relevant, but the absence of clear scope restriction, redaction, consent/warning, or sensitive-data handling makes unintended data exposure a real privacy and security risk.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
84% confidence
Finding

All user-facing instructions and examples are written in Chinese, and the document does not indicate that the skill is region-specific or that other language options are available. Per the policy, forcing a specific language without user opt-in can be a natural-language policy violation unless the locale constraint is explicitly justified.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
84% confidence
Finding

The skill description, triggers, workflow, examples, and output are presented entirely in Chinese, with no indication that users may choose another language. Under the stated policy, forcing a specific language without user opt-in is a natural-language policy concern unless the locale restriction is explicitly justified.

Content

No source excerpt is available for this finding.