Your AI Model Is A Program. Start Implementing It Like One

In early 2025, researchers discovered malicious code embedded in machine learning models on Hugging Face, exploiting model file execution to bypass security scans. This has raised awareness that AI models are executable programs, prompting new governance and security measures like model signing specifications by OpenSSF and adoption by NVIDIA. The NSA and allied agencies have issued guidance on AI supply chain risks, emphasizing the need for stricter controls on model sourcing and loading.