Back to skill

Security audit

AI全链路内容创作工厂

Security checks for vulnerabilities and agentic risk

Overview

The skill is a coherent content-production workflow, but it can reach a live Xiaohongshu publishing action without a clear final approval gate.

Review this skill before installing. It is not showing evidence of malware, credential theft, or local persistence, but it should be treated carefully because it may publish generated content to Xiaohongshu using an authenticated session. Install only if you are comfortable with that behavior, and require a final preview-and-confirm step for the exact text, images, tags, target account, and visibility before any live post.

Vulnerability Patterns
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • 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 (1)

T09 · Insecure Skill Coding Practices

Warning
Location
workflow.json:81
Finding

External Publication Proceeds Without a Final User-Approval Gate

Content
View full analysis

Vulnerability Details

File Location: workflow.json:81-89
Vulnerability Type: Missing authorization checkpoint before an external side effect
Risk Level: Medium

json
{
  "id": 5,
  "name": "Xiaohongshu publication",
  "skill": "xiaohongshu-mcp",
  "action": "publish",
  "prompt": "Rewrite the article as a Xiaohongshu image-and-text post and publish it",
  "next": "complete"
}

The snippet above is an English rendering of the original configuration values. The control fields, skill identifier, action, and transition are unchanged.

Technical Analysis

The final workflow step invokes the authenticated publish action directly and then transitions to completion. Unlike the topic-selection step, it does not define user_intervention, an approval state, a draft-only action, or any other explicit authorization checkpoint.

This configuration is inconsistent with the accompanying documentation, which states that the assistant asks whether the user wants to publish and suggests that final publication requires confirmation. Security-sensitive behavior is governed by the executable workflow configuration rather than by descriptive documentation. Consequently, an orchestrator that follows workflow.json can publish generated content immediately after the preceding steps.

The workflow also feeds externally researched material through generative transformations before publication. Inaccurate, adversarially influenced, or otherwise unsuitable source material may therefore reach an authenticated external account without the user reviewing the final text, images, tags, target account, or visibility settings.

Attack Path

  1. The search skill collects externally controlled or misleading content concerning the selected topic.
  2. The user confirms the general topic at the only explicitly configured intervention point.
  3. Downstream skills generate an article, cover image, and platform-specific adaptati ...[truncated 1242 chars]
Remediation
View remediation

Remediation Suggestions

  1. Insert a mandatory user-intervention step immediately before any external publication action.
  2. Split publication into separate prepare_draft and publish operations. The default workflow must stop after draft preparation.
  3. Present the exact final text, images, tags, target platform account, visibility, and other publication settings to the user.
  4. Require an explicit, affirmative confirmation tied to that exact payload. Editing the payload should invalidate prior approval and require confirmation again.
  5. Add a configuration control such as user_intervention: true and an unambiguous intervention label such as final_publication_approval.
  6. Fail closed when approval is absent, expired, ambiguous, or cannot be associated with the current draft.
  7. Ensure the external publication skill operates with the minimum account permissions necessary and does not reuse authorization granted for unrelated content.
  8. Add automated workflow tests verifying that execution cannot reach action: "publish" without a recorded approval event.
  9. Update the workflow and documentation together so that the configured behavior consistently matches the stated final-review requirement.
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (13)

Missing User Warnings

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The final step performs publication to 小红书 automatically and routes directly to completion without a publication-specific user confirmation. This is dangerous because generated content may be inaccurate, policy-violating, or simply unwanted, and the workflow as written can turn a drafting request into an external side effect on a live platform.

Content

No source excerpt is available for this finding.

Vague Triggers

High
Category
Not specified by scanner
Confidence
96% confidence
Finding

The trigger list is broad enough to match ordinary writing requests such as '写文章' or '帮我写文章', which can cause this workflow to activate outside a clearly content-publishing-specific context. Because the workflow chains into image generation and an eventual publishing step, accidental invocation materially increases the chance of unintended actions and misuse.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The README describes publishing to Xiaohongshu as an automated workflow step without clearly warning that this triggers an external account action using authenticated session material such as cookies. That can mislead users into approving or enabling account-impacting behavior without explicit informed consent, increasing the risk of unintended posting, data transmission to third-party services, or policy/account consequences.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The example says the assistant will automatically rewrite and publish to Xiaohongshu, but it does not disclose that this is a live external side effect affecting the user's account and public presence. In a content-publishing skill, omission of a strong safety notice is dangerous because users may interpret the action as harmless automation rather than irreversible account activity.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The trigger phrases include very broad, everyday requests such as '写文章' and '创作内容', which can cause the skill to activate in situations where the user did not intend to invoke a workflow that chains multiple tools. In this skill’s context, unintended activation is more dangerous because the workflow can automatically perform external searches, generate content, and proceed toward publication-oriented actions.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill description presents the workflow as content creation and publication, but it does not prominently warn that it may publish to an external platform via xiaohongshu-mcp. In this context, missing disclosure is risky because users may think they are only drafting content when the skill is capable of taking outward-facing actions that affect accounts, reputation, and data exposure.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
82% confidence
Finding

The skill name, description, triggers, prompts, and output style are all defined in Chinese, including platform-specific Chinese content targets such as WeChat and Xiaohongshu, with no indication that users may choose another language. Under the policy, forcing a specific language or locale without opt-in is a natural-language policy concern unless the restriction is explicitly documented and justified.

Content

No source excerpt is available for this finding.

Skill Enumeration

Medium
Category
Agent Snooping
Confidence
80% confidence
Finding

Skill enumerates or reads other installed skills. Access to other skills' SKILL.md files or the skills directory reveals prompt instructions, capabilities, and secrets that should be invisible to peer skills.

Content

Scanner excerpt · workflow.json (reported line 8)May include surrounding context.

json
"skills": [
    {
      "name": "Agent-Reach",
      "location": "workspace/skills/Agent-Reach/SKILL.md",
      "role": "热点发现",
      "step": 1,
      "inputs": ["用户指定关键词或主题"],

Skill Enumeration

Medium
Category
Agent Snooping
Confidence
80% confidence
Finding

Skill enumerates or reads other installed skills. Access to other skills' SKILL.md files or the skills directory reveals prompt instructions, capabilities, and secrets that should be invisible to peer skills.

Content

Scanner excerpt · workflow.json (reported line 16)May include surrounding context.

json
},
    {
      "name": "公众号助手",
      "location": "workspace/skills/wechat-article-pro/SKILL.md",
      "role": "文章创作",
      "step": 2,
      "inputs": ["热点话题、写作风格、关键词"],

Skill Enumeration

Medium
Category
Agent Snooping
Confidence
80% confidence
Finding

Skill enumerates or reads other installed skills. Access to other skills' SKILL.md files or the skills directory reveals prompt instructions, capabilities, and secrets that should be invisible to peer skills.

Content

Scanner excerpt · workflow.json (reported line 24)May include surrounding context.

json
},
    {
      "name": "nano-banana-pro",
      "location": "workspace/skills/nano-banana-pro/SKILL.md",
      "role": "封面/配图生成",
      "step": 3,
      "inputs": ["文章主题、封面尺寸(900x383公众号 / 3:4小红书)"],

Skill Enumeration

Medium
Category
Agent Snooping
Confidence
80% confidence
Finding

Skill enumerates or reads other installed skills. Access to other skills' SKILL.md files or the skills directory reveals prompt instructions, capabilities, and secrets that should be invisible to peer skills.

Content

Scanner excerpt · workflow.json (reported line 32)May include surrounding context.

json
},
    {
      "name": "card-renderer",
      "location": "workspace/skills/card-renderer/SKILL.md",
      "role": "知识卡片渲染",
      "step": 4,
      "inputs": ["Markdown 文案、风格选择(Mac Pro/赛博朋克/包豪斯)"],

Skill Enumeration

Medium
Category
Agent Snooping
Confidence
80% confidence
Finding

Skill enumerates or reads other installed skills. Access to other skills' SKILL.md files or the skills directory reveals prompt instructions, capabilities, and secrets that should be invisible to peer skills.

Content

Scanner excerpt · workflow.json (reported line 40)May include surrounding context.

json
},
    {
      "name": "xiaohongshu-mcp",
      "location": "workspace/skills/xiaohongshu-mcp/SKILL.md",
      "role": "小红书发布",
      "step": 5,
      "inputs": ["公众号文章改写版、配图、话题标签"],

Vague Triggers

Low
Category
Not specified by scanner
Confidence
87% confidence
Finding

The free-form invocation example ('帮我搜索今天的 AI 热点,然后写一篇公众号文章') is broad enough to normalize automatic activation from common natural-language requests. This increases the chance that a user’s ordinary content request triggers a multi-step pipeline with external tool usage before they clearly understand the scope of actions.

Content

No source excerpt is available for this finding.