中文 EN

Scan Report

Scan New Files

✓ Trusted — Risk Score 5/100
Last scan:22 hr ago Rescan
5 /100
lora-pipeline
End-to-end LoRA training pipeline: reference photo collection → face verification → dataset scraping → quality check → WD14 captioning → RunPod training
Legitimate LoRA training pipeline skill with clear documentation, enforced local-only face verification, and appropriate cloud infrastructure usage for ML training.
Skill Namelora-pipeline
Duration44.8s
Enginepi
✓
Safe to install
No action needed. The skill is well-structured with explicit privacy rules and correct use of subprocess for training.

Findings 2 items

Severity Finding Location
Low
Hardcoded user home path in tag_batch.py Doc Mismatch
Model paths in tag_batch.py are hardcoded to /Users/mini/.openclaw/workspace/tools/wd14-tagger/models/ — a code quality issue. Should use environment variables or config.
model_path = "/Users/mini/.openclaw/workspace/tools/wd14-tagger/models/model.onnx"
→ Use os.path.expanduser('~/.openclaw/...') or a configurable path.
 scripts/tag_batch.py:17
Info
No _meta.json or allowed-tools declaration Doc Mismatch
SKILL.md lacks an allowed-tools declaration block. While not a security issue, it makes capability mapping harder for automated tools.
No allowed-tools section
→ Add an allowed-tools block at top of SKILL.md matching the observed capabilities.
 SKILL.md:1
ResourceDeclaredInferredStatusEvidence
Filesystem WRITE WRITE ✓ Aligned SKILL.md: creates datasets/, scripts write outputs
Network WRITE WRITE ✓ Aligned Phase 6: runpodctl + scp to RunPod cloud
Shell WRITE WRITE ✓ Aligned batch_lora_train.py: subprocess for training; SKILL.md: scp/ssh for pod ops
Environment NONE NONE — No env var access detected
Skill Invoke NONE NONE — sessions_spawn documented, no cross-skill data theft
Clipboard NONE NONE — No clipboard access
Browser READ READ ✓ Aligned Phase 3: browser snippets for scraping (documented)
Database NONE NONE — No database access
2 findings
🔗
Medium External URL 外部 URL
https://npcnewsonline.com/.../thumb/img001.jpg
phases/02-scraping.md:50
🔗
Medium External URL 外部 URL
https://npcnewsonline.com/.../large/img001.jpg
phases/02-scraping.md:51

File Tree

9 files · 42.4 KB · 1322 lines
Markdown 6f · 871L Python 3f · 451L
├─ ▾ 📁 phases
│ ├─ 📝 01-reference.md Markdown 75L · 2.7 KB
│ ├─ 📝 02-scraping.md Markdown 68L · 2.3 KB
│ ├─ 📝 03-verify.md Markdown 223L · 7.3 KB
│ ├─ 📝 04-caption.md Markdown 87L · 2.5 KB
│ └─ 📝 05-training.md Markdown 326L · 10.0 KB
├─ ▾ 📁 scripts
│ ├─ 🐍 batch_lora_train.py Python 351L · 10.7 KB
│ ├─ 🐍 smart_crop.py Python 38L · 1.1 KB
│ └─ 🐍 tag_batch.py Python 62L · 2.2 KB
└─ 📝 SKILL.md Markdown 92L · 3.5 KB

Dependencies 5 items

PackageVersionSourceKnown VulnsNotes
deepface unpinned import No No requirements.txt; imported in phase docs for local face verification
opencv-python unpinned import No Standard CV2 for image processing
onnxruntime unpinned import No Local ONNX inference for WD14 tagger
pillow unpinned import No Image handling
numpy/pandas unpinned import No Data processing

Security Positives

✓ Phase 4 explicitly forbids cloud face verification — 'NEVER send images to any cloud API' with clear enforcement via local DeepFace
✓ Privacy rules are prominently documented: NO DATA INSPECTION, NO CLOUD UPLOAD, NO DATA LEAKAGE
✓ Subprocess usage is exclusively for ML training (Kohya sd-scripts) — expected behavior with no exfiltration risk
✓ No base64, no eval(), no obfuscation, no hidden HTML comments anywhere in the codebase
✓ Phase 6 correctly handles credential management via RunPod SSH keys — no credential harvesting
✓ Sentry monitoring is bash-only (zero LLM API calls during polling) — quota protection is well-designed
✓ Output files are .safetensors (safe format) — no arbitrary code execution from outputs
✓ No dependencies with unpinned versions — standard ML libraries (deepface, cv2, onnxruntime)