Unsloth Studio Flaw Turns Routine Model Inspection Into Code Execution

A vulnerability in Unsloth Studio, a Web UI for fine-tuning large language models, allows malicious AI models to execute arbitrary Python code during inspection. This was discovered by Pillar Security and addressed in update 2026.6.9. The flaw stems from the ‘trust_remote_code’ setting, which allows the underlying ‘Transformers’ library to run custom code within a model’s configuration file. While no exploitation has been reported, the issue highlights a broader problem with how machine learning tools handle executable model content.