Verified, then logged.
Sun 22 Nov 2026 · 00:23 KST / 15:23 UTC한국어 | ENSearch /

Breaking02 / IT

Hugging Face's transformers v5.22.2 hardens Jinja sandboxing

The patch hardens Jinja sandboxing, but the release notes do not name the specific flaw it was meant to block.

In a plainly lit server room, an engineer seen from behind updates the software on a rack of multi-GPU servers from a laptop on a rolling cart. AI-generated image

Hugging Face released transformers v5.22.2 on 6 Nov 2026, a patch for users of the library that hardens its Jinja sandboxing and corrects NCCL distributed execution, according to the release notes.

The release notes list the Jinja change as a security fix that strengthens the sandbox. Jinja is a Python templating engine, and a sandbox is a restricted environment that limits what code can do. Pull request #50431 carries the change. Its author is Rocketknight1.

On the distributed side, the patch sets device_id only when NCCL is in use. NCCL is Nvidia's library for communication between GPUs. The release notes credit michaelbenayoun and 3outeille as the authors of #50281.

The release notes were posted by vasqu. They describe the update as "Some in between patch for security and distributed fixes".

Compatibility. The change entries in the release notes carry no mark of a compatibility break. The full difference from v5.22.1 is in the changelog linked from the release notes.

The release notes do not name the specific flaw that the Jinja sandbox hardening was meant to block.