Safetensors security audit confirms safe default format

AI Impact Summary

Hugging Face, EleutherAI, and Stability AI have commissioned a security audit of the Safetensors library, confirming it as safe and ready for widespread adoption. This shift addresses critical security concerns around the use of pickle for model serialization, which could allow for malicious code execution. The library's transition to default usage will streamline model deployment and improve efficiency through features like lazy loading and framework agnosticism.