Back to skill

Security audit

纯血万相冰箱盲盒

Security checks for vulnerabilities and agentic risk

Overview

The skill does what it says, but it uploads private fridge photos and API credentials to configurable cloud endpoints without enough scoping or privacy disclosure.

Review before installing. Use only non-sensitive fridge photos, avoid visible medications or identifying labels, and do not set WAN_API_URL to an untrusted or unknown host. Prefer a dedicated DashScope API key with limited quota, run in a clean virtual environment, and be aware that the tool sends images and prompts to cloud AI services.

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

T09 · Insecure Skill Coding Practices

Error
Location
scripts/generate_gourmet.py:347
Finding

Unvalidated API endpoint can expose the bearer credential and private refrigerator images

Content
View full analysis
Remediation
View remediation

T08 · Insecure Dependencies

Note
Location
SKILL.md:139
Finding

Unpinned runtime dependency installation creates avoidable supply-chain exposure

Content
View full analysis
Remediation
View remediation
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Taint TrackingDirect Taint Flow, Variable-Mediated Taint Flow, Credential Exfiltration Chain
  • YARA SignaturesMalware Match, Webshell Match, Cryptominer Match
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Findings (21)

YARA rule 'agent_skill_mcp_tool_poisoning_metadata': MCP/tool metadata poisoning indicators in tool schemas or skill manifests [agent_skills]

High
Category
YARA Match
Confidence
80% confidence
Finding

YARA rule matched a hack tool or exploit indicator (offensive tools, reconnaissance, privilege escalation, or exploit frameworks).

Content

Scanner excerpt · scripts/generate_gourmet.py (reported line 448)May include surrounding context.

python
for word in ["汤", "soup", "炖", "stew", "braise"]):
                    cooking_method = "simmered"
                    dish_type = "soup or stew"
            
            # 构建强约束提示词
            ingredients_str = ", ".join(key_ingredients[:6]) if key_ingredients else "fresh vegetables"
            
            # 根据菜品类型添加具体的摆盘描述
            plating_description = ""
            negative_prompt = ""
            if "salad" in dish_type.lower():
                plating_description = "Fresh colorful salad in a white bowl, mixed raw vegetables with dressing, cold dish"
                negative_prompt = "NO cooked meat, NO fish, NO hot dish, NO grilled items, NO roasted meat"
            elif "roll" in dish_type.lower() or "wrap" in dish_type.lower():
                plating_description = "Elegant rolled wraps on a plate, cut to show colorful fillings inside"
                negative_prompt = "NO whole fish, NO meat slab, NO plain plate"

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The skill explicitly asks users to photograph the inside of their refrigerator and upload that image to an external AI service, but it does not warn users that the photo may contain sensitive personal information such as medications, identifiable products, household habits, or other incidental details. This creates a privacy and informed-consent gap: users may disclose more than they realize to a third-party cloud provider without understanding retention, processing, or sharing implications.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

This markdown file documents setting a WAN_API_KEY and sending a local fridge image to a remote API for generation, but it does not warn users that their image and related data may be transmitted to an external service. Under the markdown criteria for SQP-2, user-facing documentation should disclose behaviors that could affect privacy or data handling.

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 · scripts/examples.md (reported line 8)May include surrounding context.

bash
# Linux/macOS
export WAN_API_KEY="your-wan-api-key-here"
export WAN_API_URL="https://api.wan.xxx/v1/images/generate"

# Windows (PowerShell)
$env:WAN_API_KEY="your-wan-api-key-here"

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 · scripts/examples.md (reported line 12)May include surrounding context.

bash
# Linux/macOS
export WAN_API_KEY="your-wan-api-key-here"
export WAN_API_URL="https://api.wan.xxx/v1/images/generate"

# Windows (PowerShell)
$env:WAN_API_KEY="your-wan-api-key-here"

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The skill base64-encodes a user-supplied fridge photo and sends it to an external multimodal API for ingredient recognition without any explicit privacy notice or consent flow. Fridge images can reveal household habits, medications, brands, family context, or other sensitive personal information, so undisclosed transfer is a meaningful privacy risk.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
90% confidence
Finding

This is a duplicate report of the same external transmission at line 573. The underlying issue is the same: a personal fridge image is sent off-device to a cloud API without explicit warning.

Content

Scanner excerpt · scripts/generate_gourmet.py (reported line 573)May include surrounding context.

python
try:
            print(f"🔍 正在识别冰箱中的食材...")
            response = requests.post(
                "https://dashscope.aliyuncs.com/api/v1/services/aigc/multimodal-generation/generation",
                json=vl_payload,
                headers=headers,

External Transmission

Medium
Category
Data Exfiltration
Confidence
90% confidence
Finding

This is a duplicate report of the same external transmission at line 573. The underlying issue is the same: a personal fridge image is sent off-device to a cloud API without explicit warning.

Content

Scanner excerpt · scripts/generate_gourmet.py (reported line 573)May include surrounding context.

python
try:
            print(f"🔍 正在识别冰箱中的食材...")
            response = requests.post(
                "https://dashscope.aliyuncs.com/api/v1/services/aigc/multimodal-generation/generation",
                json=vl_payload,
                headers=headers,

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

Recipe generation uploads the user's fridge photo to a remote multimodal service but does not clearly tell the user that their image leaves the local environment. Because the content of a fridge photo may contain sensitive information, the hidden external transmission creates a legitimate privacy and data-handling concern.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
90% confidence
Finding

This is a duplicate report of the external transmission in the recipe-generation path at line 785. The risk remains undisclosed upload of potentially sensitive user image data to a third-party AI service.

Content

Scanner excerpt · scripts/generate_gourmet.py (reported line 785)May include surrounding context.

python
try:
            print(f"   正在生成菜谱...")
            response = requests.post(
                "https://dashscope.aliyuncs.com/api/v1/services/aigc/multimodal-generation/generation",
                json=vl_payload,
                headers=headers,

External Transmission

Medium
Category
Data Exfiltration
Confidence
90% confidence
Finding

This is a duplicate report of the external transmission in the recipe-generation path at line 785. The risk remains undisclosed upload of potentially sensitive user image data to a third-party AI service.

Content

Scanner excerpt · scripts/generate_gourmet.py (reported line 785)May include surrounding context.

python
try:
            print(f"   正在生成菜谱...")
            response = requests.post(
                "https://dashscope.aliyuncs.com/api/v1/services/aigc/multimodal-generation/generation",
                json=vl_payload,
                headers=headers,

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The surrounding comments say this mode is '基于菜谱生成' and '更贴合菜谱', but the implementation hard-codes '强制所有菜都显示为热菜' and describes the output as a hot stir-fried dish even if the parsed recipe indicates soup, salad, wraps, or other preparations. This is an active contradiction between documented intent and actual behavior.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
96% confidence
Finding

This request sends generation payloads, including user image data in some modes, to a remote service endpoint. In context, the danger is not arbitrary exfiltration to an unknown host but failure to clearly disclose third-party processing of personal content.

Content

Scanner excerpt · scripts/generate_gourmet.py (reported line 1117)May include surrounding context.

python
print(f"   时间: {time_name}")
        
        try:
            response = requests.post(
                self.api_url,
                json=payload,
                headers=headers,

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The image/video generation path sends user image data and prompts to a remote API endpoint, potentially including custom requirements and recipe-derived content, without an explicit privacy warning. In this skill context, users may reasonably assume local processing of a personal fridge photo, so silent upload increases privacy exposure.

Content

No source excerpt is available for this finding.

Tainted flow: 'query_url' from requests.post (line 1176, network input) → requests.get (network output)

Medium
Category
Data Flow
Confidence
65% confidence
Finding

Data from a source is assigned to a variable that is later passed to a sink, creating a variable-mediated taint flow.

Content

Scanner excerpt · scripts/generate_gourmet.py (reported line 1182)May include surrounding context.

python
time.sleep(poll_interval)
                print(f"   第 {i+1} 次查询...", end=" ")
                
                query_response = requests.get(query_url, headers=headers, timeout=30)
                
                if query_response.status_code != 200:
                    print(f"查询失败: {query_response.status_code}")

Tainted flow: 'query_url' from requests.post (line 1176, network input) → requests.get (network output)

Medium
Category
Data Flow
Confidence
65% confidence
Finding

Data from a source is assigned to a variable that is later passed to a sink, creating a variable-mediated taint flow.

Content

Scanner excerpt · scripts/generate_gourmet.py (reported line 1250)May include surrounding context.

python
query_url = f"https://dashscope.aliyuncs.com/api/v1/tasks/{task_id}"
        
        try:
            response = requests.get(query_url, headers=headers, timeout=30)
            
            if response.status_code == 200:
                return response.json()

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The printed message states the generated image/video was '基于您的冰箱照片独家生成', yet the code explicitly disables image use in the 2.7-t2i and 2.1-t2v branches by setting text-only payloads and even comments that these modes do not use the fridge image. This is a direct contradiction in inline/user-facing documentation about what data the model actually used.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
92% confidence
Finding

This code file contains natural-language statements and data entirely in Chinese, including the top-level description asserting a Chinese recipe database, without any indication that the language is optional or tied to a documented locale-specific requirement. Under the policy rule for language/locale constraints, forcing a specific language without user opt-in is a natural-language policy concern.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
81% confidence
Finding

The file content, headings, and inline guidance are presented entirely in Chinese, which effectively forces a specific language for users without offering an alternative or documenting a justified locale restriction. This matches the policy category for language or locale constraints expressed through natural-language content.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
91% confidence
Finding

The ingredient-identification prompt explicitly requires output '使用中文', and similar recipe prompts also require Chinese output. This is a natural-language locale constraint imposed by the skill without any visible user opt-in or configurable language selection.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
92% confidence
Finding

Both recipe prompt branches instruct the model to respond '使用中文', which enforces a specific language regardless of user preference. This is a locale-policy issue because the file contains no mechanism for choosing another language or explicitly opting into Chinese-only behavior.

Content

No source excerpt is available for this finding.