Hugging Face has announced major advancements in its Kernels project, introducing a dedicated repository type on the Hugging Face Hub to streamline the packaging, distribution, and consumption of custom AI kernels. This update establishes kernels as first-class citizens within the Hub ecosystem, offering users enhanced discoverability and detailed metadata regarding supported accelerators, operating systems, and backend versions. Security remains a central pillar of the Kernels initiative. Given that kernels execute native code with privileges equivalent to the loading Python process, Hugging Face has implemented a multi-layered defense strategy. The project now enforces a trusted publisher model; by default, the kernels package loads only kernels from organizations vetted by the community. Users may opt to load kernels from untrusted sources by explicitly passing the trust_remote_code argument during initialization. Publishers must request access through account settings to join the trusted list. Additional security measures include code signing via Sigstore's cosign. Kernels are signed using ephemeral private keys, ensuring that even if keys are leaked, their utility is limited to a short window. Signatures are validated against trusted GitHub workflows, and the kernels verify-signature tool allows developers to audit kernel integrity. While signature verification upon loading is currently in testing, the infrastructure is integrated into kernel-builder. Reproducibility is further guaranteed through Nix-based hermetic builds and the embedding of source Git SHAs within the kernels. The project has also overhauled its command-line interfaces to improve separation of concerns. The kernels library now focuses exclusively on loading and preparing kernels, while kernel-builder handles construction, resulting in leaner, more specialized tools. Environment setup is simplified via one-click installation scripts and Terraform configuration guides. A significant focus is placed on supporting agentic kernel development. kernel-builder provides a structured workflow optimized for AI agents, enabling them to scaffold, build, benchmark, and iteratively optimize kernels. Backend-specific skills assist agents in navigating hardware idiosyncrasies, while tight integration with Hugging Face Jobs facilitates automated benchmarking across diverse accelerator families. Agents can compare performance against baselines to drive optimization cycles. User experience improvements include system cards that document kernel interfaces upon push. Developers can programmatically verify compatibility using has_kernel() and diagnose issues with get_kernel_variants(), which returns detailed reasons for rejection based on CPU, OS, or framework mismatches. Technical fixes address longstanding stability concerns. Hugging Face resolved segmentation faults and data corruption issues in manylinux_2_28 builds by switching from static to dynamic linking of libstdc++. This change prevents conflicts arising from global initialization routines shared between statically linked kernels and dynamically loaded libraries like PyTorch. Kernels are now compiled using the official manylinux toolchain to maintain broad compatibility. The Kernels project aims to serve both developers and end-users, fostering a secure and efficient ecosystem for custom kernel adoption. Hugging Face invites community contributions and feedback to further refine these capabilities.