The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides instructions for executing system commands, including package management (pip install), starting web interfaces (mlflow ui), serving models locally (mlflow models serve), and programmatically retrieving version control metadata via subprocess.check_output for lineage tracking.
[INDIRECT_PROMPT_INJECTION]: The skill exposes an attack surface for indirect prompt injection by ingesting data from external MLflow tracking servers and model registries.
Ingestion points: External data is retrieved via MlflowClient search methods and mlflow.pyfunc.load_model (documented in SKILL.md and references/model-registry.md).
Boundary markers: The provided code examples do not include explicit delimiters or instructions to ignore potential injections within experiment metadata.
Capability inventory: The skill includes capabilities to execute shell commands, perform network requests via requests and boto3, and write files to the local file system.
Sanitization: The code samples do not demonstrate sanitization or validation of metadata retrieved from the tracking server.

mlflow