安全决策报告

melo-tts-metadata-creator

Skill silently installs multiple undeclared Python packages (torch, whisper, audio-separator, librosa) and executes shell commands (nvidia-smi, pip install) without disclosure in SKILL.md documentation.

安装决策优先 来源: 手动上传 扫描时间: 2026/4/4
文件 8
IOC 3
越权项 5
发现 5
最直接的威胁证据

为什么得出这个结论

1/4 个维度触发
阻止
声明与实际能力

发现 5 项声明之外的能力或越权行为。

复核
隐藏执行与外联

提取到 3 个一般风险产物,需要结合上下文判断。

通过
攻击链与高危发现

没有形成明确的恶意路径。

复核
依赖与供应链卫生

发现 7 项需要关注的依赖或供应链线索。

风险分是怎么被拉高的

Doc deception - undeclared package installations +20

SKILL.md declares only 'python' as required but silently installs torch, torchaudio, whisper, audio-separator, librosa, pydub, huggingface-hub

Doc deception - undeclared subprocess execution +15

SKILL.md does not mention subprocess/shell usage. env_manager.py runs nvidia-smi and pip install commands

Undeclared network access +10

Downloads PyTorch wheels from pytorch.org and packages from pypi.tuna.tsinghua.edu.cn not mentioned in docs

Undeclared virtual environment creation +5

Creates venv directory and restarts script in venv without disclosure

最关键的证据

中危 文档欺骗

Undeclared silent package installation

SKILL.md metadata declares only 'python' as required bin, but generate_metadata_list.py:18-20 silently installs openai-whisper, torch, torchaudio via ensure_package.pip() without user consent or documentation.

scripts/generate_metadata_list.py:18
Document all required packages explicitly in SKILL.md or move installation to a documented pre-install step.
中危 文档欺骗

Undeclared subprocess shell execution

env_manager.py executes multiple shell commands via subprocess (nvidia-smi, pip install) to set up virtual environment and install dependencies. This is essential functionality not mentioned in SKILL.md.

scripts/env_manager.py:79
Document subprocess/shell usage and the virtual environment setup process in SKILL.md.
中危 文档欺骗

Undeclared extensive package dependencies

env_manager.py lines 188-194 install audio-separator, librosa, pydub, huggingface-hub[tqdm] - 4+ additional packages not declared anywhere.

scripts/env_manager.py:188
Document all dependencies or remove unnecessary packages.
低危 供应链

Unpinned package versions

All pip installs in ensure_package.py and env_manager.py use unpinned versions (*). This can lead to supply chain attacks via dependency confusion or typosquatting on future updates.

scripts/ensure_package.py:45
Pin package versions to known-good hashes or specific versions.
低危 文档欺骗

Undeclared network access for model downloads

Whisper model downloads to ./models/ directory (generate_metadata_list.py:47) and PyTorch wheels from external URLs (env_manager.py:148,161) are not disclosed.

scripts/generate_metadata_list.py:47
Document model download behavior and network usage.

声明能力 vs 实际能力

文件系统 阻止
声明 READ
推断 WRITE
scripts/config.py:20 creates LOG_DIR.parent.mkdir
命令执行 阻止
声明 NONE
推断 WRITE
scripts/env_manager.py:79 subprocess.check_call pip install
网络访问 阻止
声明 NONE
推断 READ
scripts/env_manager.py:148 downloads from pytorch.org
环境变量 阻止
声明 NONE
推断 READ
scripts/env_manager.py:45 reads os.getenv(RUNNING_IN_VENV)
技能调用 阻止
声明 NONE
推断 WRITE
scripts/generate_metadata_list.py:18 ensure_package.pip() auto-installs

可疑产物与外联

中危 外部 URL
https://pypi.tuna.tsinghua.edu.cn/simple

scripts/ensure_package.py:50

中危 外部 URL
https://download.pytorch.org/whl/cpu

scripts/env_manager.py:148

中危 外部 URL
https://download.pytorch.org/whl/

scripts/env_manager.py:161

依赖与供应链

包名版本来源漏洞备注
openai-whisper * pip (auto-installed) Silent installation at runtime, not declared in docs
torch * pip (auto-installed) Downloaded from pytorch.org
torchaudio * pip (auto-installed) Downloaded from pytorch.org
audio-separator * pip (auto-installed) Installed in env_manager.py but not required for core functionality
librosa * pip (auto-installed) Installed but not imported anywhere in the codebase
pydub * pip (auto-installed) Installed but not imported anywhere in the codebase
huggingface-hub * pip (auto-installed) Installed but not imported anywhere in the codebase

文件构成

8 个文件 · 861 行
Python 5 个文件 · 567 行Markdown 2 个文件 · 254 行Ignore 1 个文件 · 40 行
需关注文件 · 3
scripts/env_manager.py Python · 235 行
Undeclared subprocess shell execution · Undeclared extensive package dependencies · https://download.pytorch.org/whl/cpu · https://download.pytorch.org/whl/
scripts/generate_metadata_list.py Python · 152 行
Undeclared silent package installation · Undeclared network access for model downloads
scripts/ensure_package.py Python · 73 行
Unpinned package versions · https://pypi.tuna.tsinghua.edu.cn/simple
其他文件 · README.md · logger_manager.py · SKILL.md · config.py · .gitignore

安全亮点

No evidence of credential theft or credential harvesting
No reverse shell or C2 communication patterns detected
No base64-encoded or obfuscated payloads
No access to sensitive paths like ~/.ssh, ~/.aws, or .env files
No malicious persistence mechanisms (cron, startup hooks, backdoors)
Package installation uses reputable sources (pytorch.org, pypi.tuna.tsinghua.edu.cn)
Whisper transcription logic is legitimate and matches stated purpose
Virtual environment creation is a reasonable security practice for dependency isolation