中文 EN

Scan Report

Scan New Files

⚠ Low Risk — Risk Score 25/100
Last scan:1 day ago Rescan
25 /100
turbo-whisper-local-stt
本地高性能音频转文本工具，使用 Faster-Whisper large-v3-ct2 模型
A legitimate audio transcription skill with documented ML tool behavior (venv setup, pip install, model downloads). Minor supply chain and documentation gaps exist but no malicious behavior detected.
Skill Nameturbo-whisper-local-stt
Duration44.9s
Enginepi
✓
Safe to install
Consider pinning dependency versions in ensure_package.py for reproducible builds. Add explicit declaration of subprocess usage for venv/pip operations in SKILL.md.

Findings 4 items

Severity Finding Location
Medium
Unpinned Python dependencies Supply Chain
ensure_package.py and env_manager.py install packages (torch, faster_whisper, librosa) without version constraints, risking supply chain attacks from dependency confusion or compromised packages.
subprocess.check_call([sys.executable, '-m', 'pip', 'install', '--upgrade', pip_pkg, ...])
→ Pin versions like 'torch>=2.0.0' or use a requirements.txt with hashes
 scripts/ensure_package.py:31
Medium
Undeclared subprocess execution Doc Mismatch
SKILL.md declares 'bins: python' but doesn't mention subprocess usage for venv creation, pip installation, or GPU detection via nvidia-smi. These operations run shell commands with elevated implications.
subprocess.check_call([str(venv_python), '-m', 'pip', 'install', '--upgrade', 'pip'])
→ Document shell:WRITE permission requirement and list all subprocess operations
 scripts/env_manager.py:85
Low
Third-party model repository Supply Chain
Models downloaded from user-controlled HuggingFace repository (wangminrui2022/) rather than official Faster-Whisper org. While legitimate, these are unaudited weights.
"wangminrui2022/faster-whisper-large-v3-ct2"
→ Consider verifying model checksums or using official repositories
 scripts/transcribe.py:34
Low
Unrestricted pip index-url Priv Escalation
PyTorch downloaded from https://download.pytorch.org/whl/{wheel} which is legitimate but uses direct network fetch without integrity verification.
index_url = f"https://download.pytorch.org/whl/{wheel}"
→ Consider pinning torch version with hash verification
 scripts/env_manager.py:109
ResourceDeclaredInferredStatusEvidence
Filesystem READ,WRITE WRITE ✓ Aligned SKILL.md declares audio/text file I/O; transcribe.py creates dirs/writes output
Network READ READ ✓ Aligned Downloads models from HuggingFace Hub (snapshot_download)
Shell NONE WRITE ✓ Aligned env_manager.py uses subprocess for pip install, nvidia-smi detection
Environment NONE READ ✓ Aligned env_manager.py reads/writes RUNNING_IN_VENV env var
7 findings
🔗
Medium External URL 外部 URL
https://huggingface.co/wangminrui2022/faster-whisper-base-ct2
README.md:23
🔗
Medium External URL 外部 URL
https://huggingface.co/wangminrui2022/faster-whisper-large-v3-ct2
README.md:24
🔗
Medium External URL 外部 URL
https://huggingface.co/deepdml/faster-whisper-large-v3-turbo-ct2
README.md:25
🔗
Medium External URL 外部 URL
https://pypi.tuna.tsinghua.edu.cn/simple
scripts/ensure_package.py:50
🔗
Medium External URL 外部 URL
https://download.pytorch.org/whl/cpu
scripts/env_manager.py:148
🔗
Medium External URL 外部 URL
https://download.pytorch.org/whl/
scripts/env_manager.py:161
🔗
Medium External URL 外部 URL
https://huggingface.co/
scripts/transcribe.py:76

File Tree

8 files · 38.0 KB · 855 lines
Python 5f · 682L Markdown 2f · 133L Ignore 1f · 40L
├─ ▾ 📁 scripts
│ ├─ 🐍 config.py Python 52L · 2.1 KB
│ ├─ 🐍 ensure_package.py Python 73L · 2.8 KB
│ ├─ 🐍 env_manager.py Python 235L · 10.3 KB
│ ├─ 🐍 logger_manager.py Python 55L · 2.6 KB
│ └─ 🐍 transcribe.py Python 267L · 12.0 KB
├─ 📄 .gitignore Ignore 40L · 359 B
├─ 📝 README.md Markdown 99L · 5.9 KB
└─ 📝 SKILL.md Markdown 34L · 1.9 KB

Dependencies 4 items

PackageVersionSourceKnown VulnsNotes
torch * PyPI/direct No Version not pinned, fetched from pytorch.org
faster_whisper * pip No No version constraint specified
librosa * pip No Version not pinned
huggingface_hub * pip No Version not pinned

Security Positives

✓ No credential harvesting - skill doesn't access SSH, AWS, or API keys
✓ No base64/encoded payloads or obfuscated code execution
✓ No data exfiltration - transcription output stays local
✓ No persistence mechanisms (no cron, startup hooks, or backdoors)
✓ Legitimate ML transcription functionality clearly matches stated purpose
✓ Uses virtualenv isolation to contain dependencies
✓ No reverse shells, C2 communication, or suspicious network targets