Hugging Face 🤗 Kernels — Major Updates: Security, CLI, and Agentic Development
Hugging Face announced major Kernels updates with new repo type, enhanced security with trusted publishers and code signing, and agentic development foundation.
Hugging Face 🤗 Kernels — Major Updates: Security, CLI, and Agentic Development
Hugging Face announced major updates to the 🤗 Kernels project, which aims to standardize how custom kernels are packaged, distributed, and consumed within the machine learning ecosystem. These updates include fundamental security improvements, Nix-based reproducible builds, revamped CLIs, broader framework and backend support, and a foundation for agentic kernel development.
New Repository Type: kernel
A new repository type called kernel has appeared on the Hugging Face Hub. Developers can now upload, share, and manage kernels as easily as models, datasets, and Spaces. Each kernel page shows which accelerators (GPU, TPU), operating systems, and backend versions are supported. Browse all kernels at huggingface.co/kernels.
Security Improvements
The platform introduced trusted publishers — only kernels from trusted organizations are loaded by default. Users must explicitly opt in with trust_remote_code=True for untrusted kernels. Code signing is implemented via Sigstore's cosign with ephemeral private keys. The source Git SHA1 is embedded in each kernel for provenance.
Nix-based Reproducible Builds
The project uses Nix to ensure hermetic, reproducible builds. Kernels can be recompiled from source and verified against published versions, providing a strong guarantee of integrity.
Foundation for Agentic Development
The platform is preparing for automatic kernel optimization and agent-driven development workflows, where AI agents could automatically tune and optimize kernels for specific hardware.