The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill provides an example configuration for an MLflow tracking server that includes a hardcoded PostgreSQL connection string containing plaintext credentials (postgresql://user:pass@localhost:5432/mlflow).
[COMMAND_EXECUTION]: The skill documentation includes several shell commands that interact with the system and network infrastructure, specifically pip install mlflow, mlflow ui, and mlflow models serve.
[REMOTE_CODE_EXECUTION]: The Python examples use mlflow.pyfunc.load_model and mlflow.sklearn.log_model, which facilitate the loading of serialized Python objects. This introduces a risk of arbitrary code execution if models are loaded from compromised or untrusted sources due to the underlying use of serialization libraries like pickle.
[PROMPT_INJECTION]: The skill's model registry and serving capabilities create an attack surface for indirect prompt injection.
Ingestion points: Loading models from the registry via mlflow.pyfunc.load_model in model_registry.py.
Boundary markers: None; there are no instructions provided to the agent to verify the integrity or origin of the models.
Capability inventory: The skill enables hosting local REST API services via mlflow models serve and file writes via mlflow.log_artifact.
Sanitization: None; the skill does not perform validation or security scanning of model content.

mlflow