The Agent Skills Directory

[DYNAMIC_EXECUTION]: The documentation provides examples of model persistence using the pickle and joblib libraries.
Evidence: Found in references/model_evaluation.md and references/quick_reference.md.
These methods are standard in the machine learning ecosystem but involve unsafe deserialization, which could lead to arbitrary code execution if an agent or user is induced to load a malicious model file from an untrusted source.
[EXTERNAL_DOWNLOADS]: The skill provides instructions to install several well-known and trusted Python packages.
Evidence: SKILL.md and references/quick_reference.md suggest installing scikit-learn, pandas, numpy, matplotlib, and seaborn.
Other specialized libraries like imbalanced-learn, category-encoders, and umap-learn are mentioned in the reference documentation.
These downloads target official package registries and well-known open-source projects.
[INDIRECT_PROMPT_INJECTION]: The skill documents workflows for ingesting and processing external data, creating an attack surface for indirect prompt injection.
Ingestion points: Tabular data loading is documented in SKILL.md and references/quick_reference.md using pd.read_csv('data.csv').
Boundary markers: No specific delimiters or "ignore embedded instructions" warnings are suggested for data processing.
Capability inventory: The skill includes scripts like scripts/clustering_analysis.py that perform local file writing (saving PNG plots).
Sanitization: No explicit sanitization or validation logic for external data content is provided in the documentation.

scikit-learn