The MLX inference backend uses the MLX-LM library, which imports and executes arbitrary Python files from model directories via the `model file` configuration field in the `config.json` file. This occurs because MLX-LM utilizes `importlib` to load and execute the file without a safety check or trust gate. Since the MLX backend operates without sandboxing, this leads to arbitrary code execution on the Docker host as the Docker Desktop user. An attacker can trigger this by calling the 'model-runner.docker.internal' API to pull a malicious model from a controlled OCI registry and requesting inference.