GPU VulnDB

Database/AI/ML frameworks & serving

Crawl4AI Docker API: attacker-chosen HuggingFace model lets a client run Python on the crawler node

CVSS 7.7CVE-2026-119120AI/ML frameworks & servingcurated

Impact

The Docker API server does not filter the CosineStrategy model_name field in request bodies, so an authenticated client picks which HuggingFace repository the server downloads and loads. A repository holding a pickle-format weight file executes arbitrary Python inside the server process on load when PyTorch is older than 2.6. On a GPU node that is code execution with whatever privileges and GPU access the crawler container holds, including its credentials, mounted volumes and network position. Even where torch is new enough to block the pickle path, the same field lets a client pull arbitrary large models and exhaust disk and host memory on the node.

Who can reach it

Any client authenticated to the Crawl4AI Docker API (/crawl and related endpoints) that can submit an extraction strategy body. Requires low privileges, not administrative access, and outbound network reachability from the node to HuggingFace.

What to do

Upgrade Crawl4AI to 0.9.5 or later and restart the API container. Where an immediate upgrade is not possible, pin PyTorch to 2.6 or newer so pickle weight loading is refused, block the container's outbound access to model hubs, and put the API behind an authenticated proxy that rejects client-supplied model_name. Container image replacement and restart; no node reboot.

References

Related entries

All AI/ML frameworks & serving entries

This entry is curated: imported from vendor advisories with machine assistance, not yet individually verified. Confirm against your vendor's advisory before acting, and report anything wrong.