This document outlines our security practices and procedures for the WebUI Ecosystem Trainer project.

• Code execution vulnerabilities: Arbitrary command execution prevention in API endpoints
• Malicious model downloads: File validation and verification for HuggingFace/Civitai downloads
• Data exfiltration: Protection of training datasets and API credentials
• Dependency integrity: Package supply chain security (npm + pip)
• File system access: Proper permission boundaries and path validation
• API security: Input validation on all FastAPI endpoints
• WebSocket security: Protection of real-time training log streams

• Theoretical attacks requiring physical machine access
• Vulnerabilities requiring manual code modification to exploit
• Issues affecting users who intentionally bypass safety mechanisms
• Local denial of service attacks against development server instances

1. Private Reporting: Use GitHub Security Advisories for confidential reporting
2. Report Details: Include vulnerability description, reproduction steps, and potential impact assessment
3. Response Timeline: Initial response within 48 hours, resolution target within 7 days
4. Public Disclosure: Coordinated disclosure after fix implementation and testing

• Provide clear reproduction steps and environment details
• Include potential impact assessment and exploitation scenarios
• Submit uncertain findings - we prefer false positives over missed vulnerabilities

• Input Validation: Pydantic models with strict typing on all API endpoints; Zod schemas for frontend validation
• Path Traversal Prevention: Sanitization of file paths in file browser and dataset upload endpoints
• Safe Defaults: No automatic script execution, minimal privilege requirements
• Dependency Management: Version pinning via package-lock.json (npm) and requirements.txt (pip), with Dependabot monitoring
• Error Handling: Information disclosure prevention in error messages

• URL Validation: Verification of download sources (HuggingFace Hub API, Civitai API)
• File Integrity: Extension and type verification for downloaded models and VAEs
• Size Limiting: Prevention of excessive resource consumption
• External Scanning: Integration recommendations for antivirus validation

• No Logging: API keys and tokens (HuggingFace, Civitai, W&B) excluded from log files and console output
• Client-Side Storage: Tokens stored in browser localStorage, never persisted server-side
• Minimal Permissions: Least-privilege principle for API access scopes

• Frontend: Next.js 15 (React 19) on port 3000 — serves the web UI
• Backend API: FastAPI (Python) on port 8000 — handles all business logic
• WebSockets: Real-time log streaming for training, tagging, and captioning jobs
• VastAI Deployment: Ports mapped to 13000 (frontend) and 18000 (API) behind VastAI proxy

The system downloads and executes content from external sources:

• HuggingFace Models: Community-contributed model files via huggingface_hub API
• Civitai Models: Community model repository via Civitai REST API
• PyPI Packages: Python package dependencies via pip
• npm Packages: Frontend dependencies via npm registry

• Training Scripts: GPU-intensive Python script execution via vendored Kohya SS backend
• Tagging/Captioning: ONNX model inference for WD14 tagging and BLIP/GIT captioning
• System Commands: File extraction, model downloads, and training subprocess management

• Deploy on development/training systems, not production infrastructure
• Implement appropriate network isolation and firewall rules
• Maintain current system patches and Python versions
• Verify download sources and model authenticity
• Use multi-factor authentication for external service accounts

Environment Setup:

# Use virtual environments for the Python backend
python -m venv lora_training
source lora_training/bin/activate

# Verify user permissions
whoami  # Should not return privileged user

# Maintain system updates
sudo apt update && sudo apt upgrade  # Linux

Credential Management:

# Environment variable storage
export HF_TOKEN="your_token_here"
export CIVITAI_TOKEN="your_token_here"

# Local environment files (excluded from version control)
echo "HF_TOKEN=your_token_here" >> .env

Network Security:

• Use HTTPS for all external communications
• Consider VPN usage on untrusted networks
• On VastAI, be aware that proxy URLs are publicly accessible — do not leave instances running unattended with sensitive data
• For local development, the API and frontend are bound to localhost by default

File System Management:

• Use dedicated directories for training activities
• Implement regular backup procedures for valuable training outputs
• Monitor and clean temporary file accumulation
• Download completed LoRA outputs from VastAI instances before terminating them

Security updates are provided for:

• Current Release: Full security support and immediate updates
• Previous Release: Critical security fixes for 90 days
• Legacy Versions: No security support - upgrade recommended

• Malicious model files exploiting training processes
• Network-based attacks during file downloads
• Local privilege escalation through file operations
• Information disclosure through application logs
• Path traversal via file browser and upload endpoints

• Advanced persistent threats with significant resources
• Physical access attacks on local systems
• Social engineering targeting user credentials
• Theoretical future cryptographic vulnerabilities
• Malicious actors with direct access to VastAI instance shell

• Enhanced input sanitization across all API endpoints
• Pydantic validation with strict Literal types on all route models
• Automated dependency vulnerability scanning via Dependabot

• Cryptographic verification for downloaded model files
• Process sandboxing for training execution
• Authentication layer for multi-user deployments
• Rate limiting on API endpoints

• Security Issues: GitHub Security Advisories (preferred)
• General Questions: GitHub Issues

Security Philosophy: We aim to balance practical security with usability for machine learning workflows. Perfect security is unattainable, but we strive for robust protection against realistic threat scenarios.

Last Updated: March 2026