The MLX inference backend in Docker Model Runner on macOS uses the MLX-LM library, which unconditionally imports and executes arbitrary Python files from model directories via the modelfile configuration field in config.json. When a model's config.json specifies a modelfile pointing to a Python file, MLX-LM uses importlib to load and execute it with no trustremotecode gate or equivalent safety check. The MLX backend runs without sandboxing, resulting in arbitrary code execution on the Docker host as the Docker Desktop user.

Any container on the Docker network can trigger this by calling the model-runner.docker.internal API to pull a malicious model from an attacker-controlled OCI registry and request inference.

Docker Model Runner (DMR) is software used to manage, run, and deploy AI models using Docker. Versions prior to 1.0.16 expose a POST /engines/configure endpoint that accepts arbitrary runtime flags without authentication. These flags are passed directly to the underlying inference server (llama.cpp). By injecting the --log-file flag, an attacker with network access to the Model Runner API can write or overwrite arbitrary files accessible to the Model Runner process. When bundled with Docker Desktop (where Model Runner is enabled by default since version 4.46.0), it is reachable from any default container at model-runner.docker.internal without authentication. In this context, the file overwrite can target the Docker Desktop VM disk (Docker.raw ), resulting in the destruction of all containers, images, volumes, and build history. However, in specific configurations and with user interaction, it is possible to convert this vulnerability in a container escape. The issue is fixed in Docker Model Runner 1.0.16. Docker Desktop users should update to 4.61.0 or later, which includes the fixed Model Runner. A workaround is available. For Docker Desktop users, enabling Enhanced Container Isolation (ECI) blocks container access to Model Runner, preventing exploitation. However, if the Docker Model Runner is exposed to localhost over TCP in specific configurations, the vulnerability is still exploitable.