Security researchers have disclosed a patched vulnerability in Unsloth Studio, a tool used for fine-tuning and inspecting AI models. The flaw allowed a maliciously crafted model to execute arbitrary Python code on a user's machine during routine inspection, rather than simply being analyzed as data.
The issue stems from the trust_remote_code setting, which is designed to permit code execution when loading models. In Unsloth Studio, this setting could be abused by a crafted model file, turning what should be a safe review process into a code execution risk. The attack would be triggered when a user inspects or loads the model.
According to Dark Reading, the vulnerability has been patched. Users of Unsloth Studio are advised to update to the latest version to avoid exposure. The report does not indicate active exploitation in the wild, but the risk is significant for anyone who routinely inspects untrusted models.