The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external machine learning data and model objects, which presents an inherent surface for indirect prompt injection attacks. This is documented as a risk factor due to the inclusion of workflows that write files and interact with network-based monitoring tools.
Ingestion points: Untrusted data enters the agent context through variables such as X_train, X_test, and model objects in code examples throughout SKILL.md and references/workflows.md.
Boundary markers: The instructions do not define delimiters or explicit "ignore embedded instructions" warnings for the processed data sets.
Capability inventory: The skill includes file system write operations via joblib.dump and plt.savefig, and network operations via mlflow for logging metrics in references/workflows.md.
Sanitization: No evidence of input validation or sanitization is present for the data being analyzed.
[EXTERNAL_DOWNLOADS]: The skill documents the installation of the shap library and its dependencies from a remote registry.
Source: The command uv pip install shap fetches content from the official Python Package Index (PyPI).
Trusted Source: The documentation references the official GitHub repository for the SHAP library, which is a well-known service in the data science community. These references are documented neutrally as standard project dependencies.

shap