Back to skill

Security audit

商家GEO推手

Security checks for vulnerabilities and agentic risk

Overview

This merchant marketing skill has a coherent goal, but it handles sensitive business data and logged-in publishing accounts with too little scoping, confirmation, and privacy protection.

Review carefully before installing. Use it for draft generation only unless you intentionally want account-level posting automation, avoid uploading unredacted licenses or personal photos, do not save platform passwords in the browser for this workflow, verify every generated statistic/testimonial before publication, and delete local JSON/Markdown outputs that contain merchant identity or contact data.

Vulnerability Patterns
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
  • 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
Findings (3)

T09 · Insecure Skill Coding Practices

Error
Location
scripts/extract_business_info.py:49
Finding

Sensitive business and identity information is stored in plaintext

Content
View full analysis

Vulnerability Details

File Location: scripts/extract_business_info.py:49-74, 220-224; scripts/generate_geo_content.py:895-911
Vulnerability Type: Plaintext storage of sensitive business and identity data
Risk Level: High

Vulnerable Code

python
result = {
    "enterprise_name": None,
    "credit_code": None,
    "legal_person": None,
    "registered_capital": None,
    "establishment_date": None,
    "business_scope": None,
    "address": None,
    "extraction_confidence": 0.0,
    "raw_text": []
}

# ...

result["raw_text"] = text_lines
result.update(self._parse_business_license(text_lines))
python
def save_to_file(self, info: dict, output_path: str):
    """Save information to a JSON file"""
    with open(output_path, 'w', encoding='utf-8') as f:
        json.dump(info, f, ensure_ascii=False, indent=2)
    print(f"Saved to: {output_path}")
python
elif args.interactive:
    config = interactive_setup()
    generator.set_config(config)

    config_path = "business_config.json"
    with open(config_path, 'w', encoding='utf-8') as f:
        json.dump(config, f, ensure_ascii=False, indent=2)

Technical Analysis

The OCR workflow extracts and retains the complete recognized text from business-license images, including legal representative names, unified social credit codes, registered addresses, business scope, and potentially unrelated text captured from the image. The entire result, including raw_text, is then written to an unencrypted JSON file.

The interactive content generator similarly writes legal representative details, telephone numbers, addresses, and website information to a predictable business_config.json file in the current working directory. Neither implementation:

  • Restricts file permissions.
  • Encrypts sensitive fields.
  • Minimizes retained OCR data.
  • Defines a retention period.
  • Deletes data after conten ...[truncated 1779 chars]
Remediation
View remediation

Remediation Suggestions

  1. Do not retain raw_text by default. Parse required fields in memory and immediately discard the full OCR response.
  2. Collect only fields required for the requested content. Make credit codes, legal representative names, registered capital, and license images opt-in.
  3. Add explicit consent describing which fields will be stored and published.
  4. Store temporary profiles in a securely created private directory with owner-only permissions, such as mode 0600 for files and 0700 for directories.
  5. Use encrypted storage when persistence is explicitly requested.
  6. Generate non-predictable filenames instead of always using business_config.json.
  7. Add automatic expiry and secure deletion after generation or publication.
  8. Implement a documented deletion command that removes profiles, OCR output, drafts, and uploaded material.
  9. Redact sensitive identifiers in logs, previews, and generated articles.
  10. Separate public merchant fields from private verification fields so verification data can never flow into promotional output accidentally.

T05 · Unauthorized Access and Privilege Escalation

Error
Location
SKILL.md:219
Finding

Browser automation can publish content without the documented final manual confirmation

Content
View full analysis

Vulnerability Details

File Location: SKILL.md:219-229, 390-397; references/browser_automation_guide.md:78-111
Vulnerability Type: Unreviewed consequential action through authenticated browser sessions
Risk Level: High

Vulnerable Code

markdown
1. Preparation
   - Use the `agent-browser` browser automation tool
   - The merchant must log in to each platform account in advance
   - Merchants are advised to use Chrome/Edge and save passwords

2. Publishing process
   - Step 1: Start browser and open the platform backend
   - Step 2: Open the content editing page
   - Step 3: Automatically fill in the title, body, and tags
   - Step 4: Automatically upload images
   - Step 5: Click the publish button
   - Step 6: Take a screenshot to confirm successful publication
bash
# Click publish
agent-browser click "text=发布"

# Take a screenshot
agent-browser screenshot

The Skill separately states that publication uses “automatic text filling plus manual confirmation,” but the operational instructions directly click the final publication control. The screenshot occurs only after the external action has completed.

Technical Analysis

Publishing is an externally visible, consequential operation. Safe browser automation should prepare a draft and stop before the final submission boundary unless the user gives explicit, current confirmation for the exact platform, account, and content.

The supplied workflow instead uses an already authenticated browser session and directly activates the publication button. It does not require:

  • A preview of the exact final title, body, images, tags, and location.
  • Confirmation immediately before each publication.
  • Validation that the active account belongs to the intended merchant.
  • Confirmation that the selected control is “publish” rather than another destructive or irreversible action.
  • Detection of platform dialogs that alter vis ...[truncated 1836 chars]
Remediation
View remediation

Remediation Suggestions

  1. Make draft generation and draft saving the default behavior.
  2. Stop automation before the final publish control on every platform.
  3. Display the exact platform, account identity, title, body, media, tags, location, and visibility settings.
  4. Require a fresh, explicit confirmation such as “Publish this exact draft to Platform X now.”
  5. Require separate confirmation for every platform; do not treat batch-generation consent as batch-publication consent.
  6. Re-snapshot the page and verify the target button immediately before any confirmed action.
  7. Prefer platform-supported draft APIs or OAuth scopes limited to draft creation where available.
  8. Remove advice to save passwords. Use dedicated work accounts and user-mediated login instead.
  9. Add dry-run and draft-only flags, with direct publication requiring an explicit opt-in flag.
  10. Record an audit event containing the approved content hash, target account, platform, timestamp, and user confirmation.
  11. Provide rollback guidance and direct links to delete or unpublish content if publication fails validation.

other

Error
Location
scripts/generate_geo_content.py:353
Finding

Content generator fabricates customer metrics, testimonials, results, and consent statements

Content
View full analysis

Vulnerability Details

File Location: scripts/generate_geo_content.py:353-361, 405-445, 481-543; references/content_templates.md:332-333
Vulnerability Type: Deceptive content generation and unsupported factual assertions
Risk Level: High

Vulnerable Code

python
enterprise_name = self.config.get("enterprise_name", "本公司")
city = self.config.get("city", "本地")
industry = self.config.get("industry", "服务")
established_years = self.config.get("established_years", "10")
python
**Company information:**
- Service area: {city}
- Team size: professional service team of XX people
- Customers served: more than 1000 customers
- Customer satisfaction: above 98%
python
**Service results:**
- Service satisfaction: 100%
- Completion time: three days ahead of schedule
- Cost savings: 15%

**Customer feedback:** "The entire process went very smoothly. The service team was very professional and is worth recommending!"
python
| Service satisfaction | 70% | 98% | Increased by 28% |
| Problem rate | 15% | 2% | Decreased by 87% |
| Cost control | 20% over budget | Saved 10% | Optimized by 30% |
python
*This case was provided by {enterprise_name}, and the customer authorized its publication.*
*For similar requirements, contact: 400-XXX-XXXX*

Technical Analysis

The generator presents fixed placeholders and invented statistics as completed factual statements rather than clearly marked draft fields. These assertions are emitted even when the merchant configuration contains no evidence supporting them.

Unsupported claims include:

  • Ten years of operation by default.
  • More than 1,000 customers served.
  • Satisfaction rates of 98% or 100%.
  • Certified staff and standardized processes.
  • Twenty-four-hour response commitments.
  • Cost savings and early completion results.
  • Named or attributed customer testimonials.
  • Customer authorization t ...[truncated 2356 chars]
Remediation
View remediation

Remediation Suggestions

  1. Replace every unsupported statistic, testimonial, credential, and performance assertion with an explicit unresolved placeholder.
  2. Never apply a factual default such as ten years of operation. Omit the statement if no verified value exists.
  3. Introduce provenance metadata for every claim, identifying the uploaded document or user statement supporting it.
  4. Require evidence-backed fields for customer counts, satisfaction rates, certifications, savings, response times, and completion results.
  5. Mark unverified claims in the preview and block publication until they are removed or verified.
  6. Require explicit confirmation that each testimonial is authentic and approved for public use.
  7. Do not state that customer authorization exists unless consent evidence was affirmatively supplied and recorded.
  8. Use conspicuous template tokens such as [REQUIRES VERIFIED CUSTOMER COUNT] rather than realistic-looking invented values.
  9. Add pre-publication validation that rejects unresolved placeholders such as XX, 20XX, generic telephone numbers, and generic customer names.
  10. Separate factual merchant data from stylistic generation so presentation changes cannot create new facts.
  11. Add automated tests confirming that sparse input cannot produce unsupported metrics, testimonials, credentials, or consent claims.
  12. Require a final merchant attestation that the complete article is accurate and authorized before enabling any publication workflow.
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Findings (25)

Tp4

High
Category
MCP Tool Poisoning
Confidence
84% confidence
Finding

The code substantially matches part of the description: it does generate GEO/SEO-style promotional content for several specified platforms and supports draft export in batch. However, the declared description claims broader capabilities that are absent from the code, especially platform content management, one-click publishing, and processing uploaded business materials such as storefront photos or business licenses. The implementation is limited to offline template-based content generation from structured config data and saving results locally. Because these missing capabilities are prominent in the declared purpose, this is a meaningful description/behavior mismatch.

Content

No source excerpt is available for this finding.

Missing User Warnings

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The skill proactively asks users to upload sensitive business and personal documents, including business licenses and personal photos, without a clear upfront privacy warning, retention policy, minimization rule, or handling safeguards. This is especially risky because these materials may contain personal identifiers, legal entity information, and images that could be exposed, misused, or repurposed across third-party platforms.

Content

No source excerpt is available for this finding.

Undeclared Tool Scope

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.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The trigger phrases are broad enough to match ordinary requests such as 'help me write an article' or 'help me promote,' which can cause the skill to activate unexpectedly and begin soliciting sensitive business documents. In this context, unintended activation is more dangerous because the skill requests营业执照, photos, and potentially initiates browser-assisted publishing workflows on external platforms.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The data-safety statement tells users their uploaded materials are only used for content generation, but the documented workflow also uses those materials for browser-driven posting and image upload to third-party platforms. This is dangerous because users may provide sensitive documents and photos under a narrower consent model than the system actually uses, creating privacy, compliance, and unauthorized-sharing risks.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

The skill claims OAuth2.0/API-key based authorization, but the operational design actually relies on browser automation against already logged-in merchant accounts. This mismatch is dangerous because it misrepresents the authentication model, obscures session/cookie risks, and may cause users to overtrust the security controls protecting account actions.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The guide explicitly recommends enabling browser password saving ('记住密码') for accounts used in automated publishing workflows. In a skill centered on browser automation across multiple social-media platforms, stored credentials increase the risk of account takeover if the browser profile, host machine, or automation environment is accessed by another user, malware, or a compromised tool. The skill context makes this more dangerous because it encourages repeated login reuse across many business accounts, concentrating credential exposure.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
71% confidence
Finding

The entire guide mandates Chinese-language interaction and includes locale-specific examples such as collecting location details in Shenzhen and generating content for Chinese platforms, but it does not state that this is an optional or region-specific mode. Without explicit opt-in or documented justification, this creates a language/locale policy concern.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The trigger conditions are broad enough to capture generic writing-help requests such as '写篇文章' without clearly limiting activation to merchant promotion contexts. This can cause the skill to collect business-oriented materials and steer users into disclosure flows when their request may be unrelated, creating unnecessary data collection and scope confusion.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The guide instructs users to upload storefront photos, product images, and especially business licenses without any privacy notice, minimization guidance, retention policy, or warning about sensitive information exposure. These materials can contain addresses, legal entity details, registration numbers, and other identifiable business data, so collecting them by default increases the risk of oversharing and downstream misuse.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The document is written as a standardized template library entirely in Chinese and presents all output structures, headings, and examples in Chinese without indicating that language selection is optional. This can violate language/locale policy when a skill constrains responses to a specific language absent user opt-in or a clearly documented regional justification.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
83% confidence
Finding

The entire skill content is written as prescriptive guidance in Chinese and does not indicate that users may choose another language or locale. Under the stated policy, forcing a specific language without opt-in can constitute a natural-language policy violation unless the locale constraint is clearly documented and justified.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

This section provides step-by-step browser automation to publish content to Baijiahao under the user's authenticated account, but it does not warn that the action can create public posts, affect the user's brand presence, or trigger irreversible external side effects. In this skill's context, the risk is heightened because the skill is explicitly designed for multi-platform promotion and may encourage routine one-click publishing, making unintended or manipulated posting more likely.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The Toutiao workflow similarly instructs the agent to open the creator backend and publish content without clearly disclosing that the post will be made through the user's external account and may become publicly visible. Because the skill targets promotional distribution across platforms, omission of that warning creates a meaningful risk of accidental publication, spammy behavior, or misuse of an already logged-in session.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The Zhihu instructions omit any notice that article or answer submission will be attributed to the user's account and may immediately affect their public profile and credibility. In a reputation-sensitive platform like Zhihu, automated posting without explicit disclosure and confirmation is especially risky because content can be perceived as authoritative and may be difficult to retract cleanly after publication.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The Xiaohongshu workflow automates note publication steps but does not warn the user that the result may be a public branded post under their account, potentially including location and promotional claims. The danger is amplified by this skill's stated purpose of marketing and content distribution, where automation can quickly spread unreviewed or policy-violating material across consumer-facing channels.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The remaining platform workflows (e.g., Douyin, Sohu, NetEase, Kuaishou) continue the same pattern: they provide operational steps for external publication without user-facing warnings about public posting, account attribution, or possible policy/account consequences. This is a true vulnerability because the skill is explicitly oriented toward broad distribution of promotional content, so silent automation materially increases the chance of unauthorized, accidental, or manipulative posting across multiple services.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

This code file contains natural-language instructions and CLI help text that force a specific language/locale experience for all users. Under the policy, locale-specific behavior should either be optional for the user or clearly justified as region-specific; neither is present here.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The embedded publishing guides and warnings are all written in Chinese, and the script provides no mechanism for users to select another language. This is a natural-language locale policy issue because the file hard-codes one language rather than offering a choice or clearly justifying the limitation.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The OCR engine is initialized with a fixed Chinese language setting, and the script's prompts, help text, and generated output are entirely Chinese. This enforces a specific language/locale without user opt-in or documented justification, which matches the policy category for language or locale constraints.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The script extracts and persists highly sensitive business and personal data from licenses, including unified credit code, legal representative name, address, and business scope, then writes it to JSON without any privacy notice, consent workflow, retention guidance, or file permission controls. In this skill context, users are explicitly encouraged to upload营业执照 and storefront materials, so the code increases the risk of accidental local disclosure, oversharing, or downstream misuse of regulated business identity data.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The script's docstrings, CLI descriptions, prompts, generated content, and platform labels are entirely hard-coded in Chinese. This enforces a single language/locale experience without any documented opt-in, fallback, or configuration for users who may require another language.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

L438 的函数注释说明该函数用于“生成客户案例内容”,但在 L478 实际插入了未定义的 service 变量。调用该分支时会触发运行错误而不是生成案例内容,因此文档化意图与实际行为相矛盾。

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

Interactive mode collects merchant identity and contact data, including legal representative name, phone, and address, then writes it to business_config.json on disk automatically without an explicit warning, consent step, or guidance on storage location and sensitivity. In the skill context, this data is business-sensitive PII/business identity material, so silent persistence increases the risk of unintended disclosure through shared machines, backups, logs, or later misuse by other tools.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
82% confidence
Finding

A language or locale policy issue applies across all file types when a skill effectively forces a single language without user opt-in. This guide presents all instructions only in Chinese and does not indicate that the skill is region-specific or provide an option for another language.

Content

No source excerpt is available for this finding.