The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data in the form of machine learning datasets, which serves as a potential attack surface for instructions embedded in feature values.
Ingestion points: Dataframes such as X_test and y_test are passed to various explainer classes in SKILL.md and references/workflows.md.
Boundary markers: No specific delimiters or "ignore" instructions are present to prevent the agent from potentially interpreting data content as instructions.
Capability inventory: The skill utilizes Bash(python:*) and Bash(uv:*) to execute code for computing SHAP values and generating visualizations.
Sanitization: There is no evidence of input validation or content filtering for the processed datasets.
[DYNAMIC_EXECUTION]: The skill implements standard model persistence patterns that utilize deserialization.
Evidence: The use of joblib.dump and joblib.load to save and restore explainer and model objects is documented in references/workflows.md.
Risk: While a standard practice in ML, joblib involves deserialization which can be unsafe if the agent is directed to load an untrusted file.
[EXTERNAL_DOWNLOADS]: The skill references the installation of common data science libraries.
Evidence: The installation section in SKILL.md uses uv add and uv run --with to fetch the shap and matplotlib packages from official registries.
[COMMAND_EXECUTION]: The skill generates and executes Python scripts to perform model analysis and visualization.
Evidence: Workflows in SKILL.md and references/workflows.md provide Python snippets intended for local execution to produce beeswarm, waterfall, and scatter plots.

alterlab-shap