CVE-2026-80047: Hugging Face Transformers downloads custom generation code before trust consent

Other code-loading paths, such as AutoConfig, AutoModel and AutoTokenizer, check consent first. load_custom_generate() fetches and caches the remote module in dynamic_module_utils.py before the check. Attacker-controlled Python from custom_generate/generate.py is copied into ~/.cache/huggingface/modules even when the user says no.

Execution itself is still gated. The harm is the file left behind. It persists across sessions and cannot be undone by declining. If a later trusted model load uses the same cache path, a stale cache collision can cause the attacker's cached code to run. That is how persistent files become code execution.

The attacker needs the victim to call load_custom_generate() against a model repo the attacker controls. They publish such a repo and get it loaded, for example through a tutorial, a shared notebook, or an agent that pulls models. The CVSS vector assumes local access with low privileges, and the advisory describes unauthorized files on disk and conditional execution, not direct remote execution.

The transformers pip package from 4.49.0 through 5.8.1. Code that never calls load_custom_generate() does not hit this path.

pip show transformers

Then look in the cache for unexpected modules:

ls ~/.cache/huggingface/modules

No patched version is listed. CERT/CC states that no vendor fix or statement existed when it published its note. A fix is proposed in pull request 48620 (commit cbc1651), but the sources do not name a release containing it, so check the release notes before relying on it.

Until then, use the workarounds from CERT/CC:

• Do not call load_custom_generate() with untrusted model repositories.
• Inspect and clear ~/.cache/huggingface/modules regularly to remove unexpected files.

List ~/.cache/huggingface/modules and look for directories or generate.py files you did not expect, especially ones that match model repos you declined to trust. Compare file times against your load_custom_generate() calls. In shared environments, check each user's cache and any cache directory set through Hugging Face environment variables.

A vulnerability in Hugging Face Transformers (versions 4.49.0, <= 5.8.1) allows remote Python files to be written to local disk without user consent when using GenerativePreTrainedModel.load_custom_generate(). The function fetches and caches a remote module file before performing the required trust_remote_code consent check, inverting the security model enforced by other code-loading paths (such as AutoConfig, AutoModel, and AutoTokenizer). As a result, attacker‑controlled Python code from custom_generate/generate.py is copied into the user’s ~/.cache/huggingface/modules directory even if the user declines the trust prompt. Although execution is correctly gated, the file write is not reversible and can persist across sessions. This can lead to persistent, unauthorized files on disk and stale cache collisions where cached attacker code may later be executed during trusted model loads. The issue stems from an unconditional file write in dynamic_module_utils.py prior to any trust verification.