The Agent Skills Directory

[COMMAND_EXECUTION]: The skill requires executing various shell commands to manage the MLflow environment and deployments.
Evidence includes commands like pip install mlflow, mlflow ui, mlflow server, mlflow models serve, mlflow deployments create, and docker run to manage models and infrastructure.
The agent is also instructed to use curl for testing local inference endpoints.
[EXTERNAL_DOWNLOADS]: The skill installs standard Python packages from the official PyPI registry.
Evidence includes instructions to install mlflow, sqlalchemy, boto3, and other ML-related libraries like scikit-learn, pytorch-lightning, and transformers as described in the dependencies and reference guides.
[INDIRECT_PROMPT_INJECTION]: The skill includes functionality to load models and artifacts from external or remote sources, which represents a potential attack surface if the sources are untrusted.
Ingestion points: Untrusted data enters the agent context through mlflow.pyfunc.load_model() and client.download_artifacts() when referencing model URIs (e.g., models:/, runs:/, or S3 paths).
Boundary markers: The instructions do not define specific delimiters or "ignore embedded instructions" warnings for the content of loaded models or logged metadata.
Capability inventory: The skill has capabilities for starting local servers (mlflow server), performing network operations for cloud deployments (mlflow sagemaker deploy), and writing artifacts to the local filesystem or cloud storage.
Sanitization: There is no mention of sanitizing or validating model signatures or artifact contents before processing them via the MLflow API.

mlflow