“The compressibility of model weights significantly heightens the exfiltration risk for large language models (LLMs).”