Secure AI kernels now
Custom AI kernels just got a security upgrade—no more malicious code slipping through. 🔒
What the video says
Hugging Face just made a decisive move to lock down custom AI kernels before a single malicious one can slip through. Their new Kernels project rolls out trusted publishers and cryptographic code signing, so every kernel runs with a verified digital fingerprint.
A new repository type on the hub now treats kernels as first-class citizens, letting users instantly see which accelerators, operating systems, and backends each one supports. System cards spell out compatibility up front, while the revamped CLI splits building from loading so agentic workflows can spin up reproducible kernels on demand.
The shift to manylinux_2_28 tooling and dynamic libstdc++ linking squashes segfaults that used to crash PyTorch processes when kernels loaded. With 2 years of Torch Stable ABI support and Apache TVMFFI in the box, The project is betting big on a future where every AI stack runs custom kernels securely.
