If you discover a security vulnerability in llm-patch, please do not open a public issue.

Instead, report it privately:

1. Email: Send details to the maintainers via the contact information in the repository.
2. GitHub Security Advisories: Use GitHub's private vulnerability reporting to submit a confidential report.

• Description of the vulnerability
• Steps to reproduce
• Potential impact
• Suggested fix (if any)

• Acknowledgment: Within 48 hours
• Initial assessment: Within 1 week
• Fix or mitigation: Depends on severity, typically within 2 weeks for critical issues

This policy covers the llm-patch Python library itself. Security issues in dependencies (PyTorch, transformers, PEFT, etc.) should be reported to those projects directly.

• Model weights: Generated LoRA adapters contain numerical weight matrices, not executable code. However, always load adapters from trusted sources.
• safetensors format: llm-patch uses the safetensors format specifically because it prevents arbitrary code execution during deserialization (unlike pickle-based formats).
• File watching: The watchdog-based file watcher only reads files from directories you explicitly configure. It does not execute file contents.
• No network access: llm-patch does not make network requests. All processing is local.