🤗 Kernels: Major Updates

2026-07-06 · Source: Hugging Face - Blog · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Cybersecurity & Data Privacy · Depth: Intermediate, medium

Summary

Hugging Face has significantly updated its 🤗 Kernels project, introducing a new "kernel" repository type on the Hub as of July 6, 2026. This new type standardizes custom kernel packaging and distribution, offering users detailed information on supported accelerators, operating systems, and backend versions, enhancing discoverability. A major focus has been on security, implementing "trusted kernel publishers" and code signing using Sigstore's cosign with ephemeral private keys to prevent malicious kernel execution and protect against compromised credentials. The project also revamped its CLIs for `kernels` and `kernel-builder` for clearer separation of concerns, extended framework support to include Torch Stable ABI and Apache TVM FFI, and laid a foundation for agentic kernel development by integrating with HF Jobs for automated benchmarking. Environment setup is simplified with an installation script and Terraform guide, and system cards provide essential kernel information.

Key takeaway

For AI Security Engineers or ML Engineers deploying custom kernels, Hugging Face's updated Kernels project significantly enhances security and reliability. You should prioritize loading kernels only from trusted publishers by default, and explicitly use `trust_remote_code=True` with extreme caution for unverified sources. Utilize the new `kernels verify-signature` tool to validate signed kernels, mitigating risks from compromised credentials. This framework streamlines secure kernel integration and development.

Key insights

The 🤗 Kernels project now offers a secure, standardized Hub for custom kernels, supporting agentic development and broad framework compatibility.

Principles

• Security for native code execution requires multi-layered defense.
• Reproducibility and provenance are critical for trusted kernel builds.
• Standardized interfaces enable cross-framework kernel interoperability.

Method

The article describes a process for agentic kernel development: scaffold, build, benchmark, and iteratively optimize kernels using `kernel-builder` and `kernels` CLIs, integrated with HF Jobs for performance evaluation across hardware.

In practice

• Use `get_kernel` with `trust_remote_code=True` for untrusted kernels.
• Verify kernel compatibility with `has_kernel()` or `get_kernel_variants()`.
• Set up `kernel-builder` with the quick install script or Terraform.

Topics

• Hugging Face Kernels
• Custom Kernel Development
• AI Security
• Code Signing
• Agentic AI
• Apache TVM FFI

Code references

• huggingface/kernels

Best for: CTO, VP of Engineering/Data, Director of AI/ML, Machine Learning Engineer, AI Engineer, AI Security Engineer

Related on AIssential

• Malicious Hugging Face Models Could Trigger Remote Code Execution — Malicious AI model configurations can exploit library processing mechanisms to achieve remote code execution, even bypassing explicit security flags.
• From AI Agents to Faster Kernels: Ben Burtenshaw & Felix LeClair (AI Plumbers #2) — AI agents can generate optimized kernels, but structured "skills" and robust infrastructure are crucial for widespread adoption and benefit.
• The third major Linux kernel flaw in two weeks has been found - thanks to AI — AI-driven tools are accelerating the discovery of critical open-source security vulnerabilities like Fragnesia.
• Article: Kernel-Level Ground Truth: Why eBPF is Replacing User-Space Agents for Security Observability — eBPF enables robust, low-overhead kernel-level security observability, immune to user-space compromise.
• 244,000 AI Developers Downloaded What They Thought Was OpenAI and It Stole Their Crypto Wallets and… — A sophisticated supply chain attack exploited trust in AI platforms by mimicking a legitimate OpenAI tool to steal sensitive user data.
• From 732 bytes to nowhere: shutting down Copy Fail in production — Linux kernel page cache vulnerabilities like Copy Fail amplify local exploits into cross-tenant risks in shared AI platforms.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Hugging Face - Blog.