IncidentMalicious model files
Unsloth Studio: selecting a model ran the code shipped with it
Pillar Security found that Unsloth Studio, the web UI of the open-source fine-tuning library, loaded a model's configuration with trust_remote_code hard-coded on. Selecting a model whose config.json pointed to custom Python files ran that code before any weights were loaded. It was reported privately in early June and fixed in 2026.6.9 on June 18; the maintainers declined to publish a CVE, and Pillar published its write-up on September 29. No exploitation in the wild was reported.
Replay the attack step by step in the Learning Center
How it happened
- A user picks a model in the Studio UI, for example from Hugging Face.
- The backend calls load_model_config() with trust_remote_code=True.
- The model's config.json maps its configuration class to Python files in the repository, which run when the configuration is instantiated.
Impact
Code running as the user on GPU development machines, with access to Hugging Face tokens, SSH keys, cloud credentials and training data.
The fix
Upgrade Unsloth to 2026.6.9 or later, and rotate tokens and keys on machines where untrusted models were selected.
Lessons
- Inspecting a model can be enough to run it: keep trust_remote_code off in every tool, including previews.
- A fix without an advisory or CVE will not reach your scanners; follow research write-ups too.