Back to skill

Security audit

AI Video Repurpose × Digital Human Dubbing

Security checks for vulnerabilities and agentic risk

Overview

The skill is a coherent AI video web app, but it exposes powerful media-processing and file-writing behavior with weak access controls and misleading privacy cleanup claims.

Install only in an isolated local environment, bind to 127.0.0.1 unless you add authentication, do not expose the service publicly, do not share browser cookies with an assistant or third party, and assume uploaded/downloaded media and transcripts may remain on disk until you manually clean them. Review copyright, consent, likeness, and privacy obligations before processing anyone else's content.

Vulnerability Patterns
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
  • 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
Findings (6)

T09 · Insecure Skill Coding Practices

Error
Location
backend/main.py:276
Finding

Server-Side Request Forgery Through Unrestricted Video URLs

Content
View full analysis
str: """Resolve a short URL to its destination.""" import requests try: resp = requests.get( url, allow_redirects=True, timeout=10, headers={"User-Agent": "Mozilla/5.0"}, ) return resp.url except Exception: return url ``` ```python process = await asyncio.create_subprocess_exec( *cmd + [url], stdout=asyncio.subprocess.PIPE, stderr=asyncio.subprocess.PIPE, ) ``` ### Technical Analysis The API validates only that the submitted string begins with `http://` or `https://`. It does not restrict destination hostnames to the supported video platforms, resolve and reject non-public IP addresses, or validate redirect destinations. The URL is subsequently passed to `yt-dlp`. In the Douyin fallback path, it may also be requested directly with `requests.get()` while redirects are enabled. Consequently, an unauthenticated caller can cause the server to initiate HTTP requests to attacker-selected destinations from the server's network context. The following destination classes are not blocked: - Loopback addresses such as `127.0.0.1` and `::1` - Private network ranges - Link-local addres ...[truncated 1301 chars]
Remediation
View remediation

T05 · Unauthorized Access and Privilege Escalation

Error
Location
backend/main.py:399
Finding

Path Traversal and Arbitrary File Overwrite Through Reference Video Filename

Content
View full analysis
Remediation
View remediation

T05 · Unauthorized Access and Privilege Escalation

Error
Location
backend/main.py:276
Finding

Missing Authentication and Task-Level Authorization on a Network-Exposed Service

Content
View full analysis
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Warning
Location
backend/main.py:399
Finding

Unbounded Reference Video Upload Can Exhaust Memory and Disk

Content
View full analysis
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Warning
Location
backend/task_manager.py:43
Finding

Sensitive Media and Transcripts Are Retained Despite Documented Cleanup

Content
View full analysis
Remediation
View remediation

T08 · Insecure Dependencies

Warning
Location
backend/requirements.txt:1
Finding

Unpinned Third-Party Dependencies Produce Non-Reproducible and Unverified Installations

Content
View full analysis
=0.104.0 uvicorn[standard]>=0.24.0 python-multipart>=0.0.6 pydantic-settings>=2.1.0 yt-dlp>=2023.10.0 openai-whisper>=20231117 openai>=1.6.0 edge-tts>=6.1.0 pydub>=0.25.1 soundfile>=0.12.1 torch>=2.1.0 torchvision>=0.16.0 gfpgan>=1.3.8 basicsr>=1.4.2 facexlib>=0.3.0 opencv-python>=4.8.0 moviepy>=1.0.3 Pillow>=10.1.0 aiofiles>=23.2.1 redis>=5.0.0 ``` ```bash pip install -r backend/requirements.txt ``` ### Technical Analysis All declared packages use open-ended lower-bound constraints. There is no lockfile, resolved transitive dependency manifest, package hash verification, or documented trusted package index. A future installation can therefore resolve substantially different direct and transitive dependency versions from those present when the Skill was audited. Python package installation may execute package build logic, and imported dependencies execute with the privileges of the application account. The audit did not identify a currently known malicious package name in this list. The issue is the absence of reproducibility and integrity controls, not evidence that a listed dependency is presently malicious. ### Attack Path 1. A user follows the setup instructions and runs `pip install -r backend/requirements.txt`. 2. Pip resolves the newest versions satisfying the broad lower bounds at installation time. 3. A direct or transitive dependency may have changed, become compromised, or introduced a vulnerable release after the Skill was reviewed. 4. The package is downloaded without verification against an audited hash set. 5. Installation or later import executes the newly resolved code with the user's privileges. ### Impact Assessment A compromised or unsafe dependency could obtain the same effective privileges as the installation or ...[truncated 383 chars]
Remediation
View remediation
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • 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 (92)

Tp4

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

This mismatch becomes security-relevant because the skill omits disclosure of subprocess invocation and external tooling despite requiring ffmpeg, yt-dlp, and service integrations. Hidden execution surfaces make it harder to assess risk and can lead operators to authorize a skill with more powerful local execution than expected.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
94% confidence
Finding

This mismatch becomes security-relevant because the skill omits disclosure of subprocess invocation and external tooling despite requiring ffmpeg, yt-dlp, and service integrations. Hidden execution surfaces make it harder to assess risk and can lead operators to authorize a skill with more powerful local execution than expected.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

This mismatch becomes security-relevant because the skill omits disclosure of subprocess invocation and external tooling despite requiring ffmpeg, yt-dlp, and service integrations. Hidden execution surfaces make it harder to assess risk and can lead operators to authorize a skill with more powerful local execution than expected.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

This mismatch becomes security-relevant because the skill omits disclosure of subprocess invocation and external tooling despite requiring ffmpeg, yt-dlp, and service integrations. Hidden execution surfaces make it harder to assess risk and can lead operators to authorize a skill with more powerful local execution than expected.

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 43)May include surrounding context.

md
pip install -r backend/requirements.txt

Credential Access

High
Category
Privilege Escalation
Confidence
60% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · backend/config.py (reported line 48)May include surrounding context.

python
PORT: int = 8080

    class Config:
        env_file = ".env"
        env_file_encoding = "utf-8"

Possible Typosquatting: 'uvicorn' resembles popular package 'gunicorn'

High
Category
Supply Chain
Confidence
70% confidence
Finding

Package name closely resembles a popular package, suggesting possible typosquatting. Attackers publish malicious packages with similar names to trick developers into installing them.

Content

No source excerpt is available for this finding.

Hidden Instructions

High
Category
Prompt Injection
Confidence
70% confidence
Finding

Hidden instructions were detected in comments or invisible text. These could contain malicious directives. Manual review is recommended.

Content

Scanner excerpt · frontend/index.html (reported line 151)May include surrounding context.

html
<div class="badge">⚡ AI Transcribe · Editable · Upload Reference Video</div>
  </div>

  <!-- Input Area -->
  <div class="input-card">
    <label for="videoUrl">🔗 Short Video URL</label>
    <div class="input-row">

Credential Access

High
Category
Privilege Escalation
Confidence
60% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · SKILL.md (reported line 46)May include surrounding context.

md
pip install -r backend/requirements.txt

# 2. 配环境(至少写 API Key)
echo 'OPENAI_API_KEY=sk-xxx' > backend/.env

# 3. 启动
cd backend && uvicorn main:app --host 0.0.0.0 --port 8080

Credential Access

High
Category
Privilege Escalation
Confidence
60% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · SKILL.md (reported line 234)May include surrounding context.

md
pip install -r backend/requirements.txt

# 2. 配环境(至少写 API Key)
echo 'OPENAI_API_KEY=sk-xxx' > backend/.env

# 3. 启动
cd backend && uvicorn main:app --host 0.0.0.0 --port 8080

Credential Access

High
Category
Privilege Escalation
Confidence
60% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · references/quickstart.md (reported line 14)May include surrounding context.

md
pip install -r backend/requirements.txt

# 2. 配环境(至少写 API Key)
echo 'OPENAI_API_KEY=sk-xxx' > backend/.env

# 3. 启动
cd backend && uvicorn main:app --host 0.0.0.0 --port 8080

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The README promotes a workflow that downloads third-party social media videos, rewrites content, and generates a digital-human video that appears to depict the user speaking, but it provides no warning about consent, likeness rights, copyright, or privacy obligations. In this context, the omission materially increases the risk of non-consensual impersonation, unauthorized reuse of content, and processing of personal data without appropriate notice or safeguards.

Content

No source excerpt is available for this finding.

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
92% confidence
Finding

The skill advertises substantial capabilities including network access, file I/O, and running local services/subprocesses, but it does not declare any explicit tool scope or permission boundaries. This increases the risk of overbroad execution in an agent platform because reviewers and runtime policy engines cannot easily constrain what the skill is allowed to do.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

Broad trigger phrases like general video-generation or talking-head terms can cause the skill to activate in contexts the user did not intend. In an agent environment, unintended invocation may route user content, files, or URLs into a workflow involving uploads, downloads, transcription, and media generation without sufficiently specific consent.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

This skill handles sensitive user media including videos, audio, photos, transcripts, and generated likeness content, yet the description lacks a prominent upfront privacy and consent warning. In this context, that omission is dangerous because users may unknowingly submit biometric/identity data or third-party content without understanding retention, transmission, and consent implications.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
60% 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 · SKILL.md (reported line 106)May include surrounding context.

Submit a task (URL)

bash
curl -X POST http://localhost:8080/api/tasks \
  -H "Content-Type: application/json" \
  -d '{
    "video_url": "https://www.douyin.com/video/xxxxx",

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
82% confidence
Finding

The module docstring is written only in Chinese ("应用配置"), which indicates a fixed language presentation without offering an alternative or user choice. The policy explicitly flags language/locale constraints when the skill forces a specific language without opt-in, and no opt-in or justification appears in this file.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The manifest describes a 5-step pipeline ending in a digital human video that looks like the user said it. However, this code's 'digital human' stage invokes ffmpeg only to replace/add audio to an existing video, loop a default avatar image, or generate a blank background video, with no visible lip-sync or avatar animation logic.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The request model sets the default TTS voice to "zh-CN-XiaoxiaoNeural", which hard-codes a Chinese locale rather than offering a user choice. Under the policy, forcing a specific language or locale without opt-in is a natural-language policy violation, and the surrounding Chinese-only descriptions reinforce that locale assumption.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

This model again defaults tts_voice to "zh-CN-XiaoxiaoNeural", imposing a Chinese locale during a later workflow step without any visible opt-in or alternative. Because the file does not document a justified region-specific restriction, this appears to violate the language/locale policy.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

This code silently creates a default image file and writes an output video file, while also invoking ffmpeg via subprocess. Although there are internal comments, there is no user-facing disclosure, confirmation, or visible warning in this file about these file-modifying and subprocess-execution behaviors.

Content

No source excerpt is available for this finding.

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
70% confidence
Finding

subprocess module calls execute external commands. Without careful input validation, this enables command injection.

Content

Scanner excerpt · backend/modules/digital_human.py (reported line 113)May include surrounding context.

python
if not default_img.exists():
            # 生成纯色背景
            default_img.parent.mkdir(parents=True, exist_ok=True)
            subprocess.run([
                "ffmpeg", "-y",
                "-f", "lavfi", "-i", "color=c=#2D3748:s=720x1080:d=1",
                "-frames:v", "1",

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
70% confidence
Finding

subprocess module calls execute external commands. Without careful input validation, this enables command injection.

Content

Scanner excerpt · backend/modules/digital_human.py (reported line 121)May include surrounding context.

python
], check=True, capture_output=True)

        # 用图片 + 音频生成视频
        subprocess.run([
            "ffmpeg", "-y",
            "-loop", "1",
            "-i", str(default_img),

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The module automatically detects and uses a cookies.txt file for authenticated downloads, but there is no explicit user consent flow, warning, or validation around handling these sensitive session credentials. In practice this can lead to silent use of privileged cookies, accidental account misuse, and increased blast radius if the host or logs are compromised.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The error message explicitly tells users they can send Douyin cookies to an assistant for handling, encouraging credential/session-token sharing. Cookies for logged-in social platforms can grant account access and can be reused or exfiltrated outside the downloader's trust boundary, making this especially dangerous in a tool that already consumes authenticated cookies for downloads.

Content

No source excerpt is available for this finding.