The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external data for machine learning modeling and statistical analysis, which creates a surface for indirect prompt injection.
Ingestion points: The skill uses pandas to read CSV and Parquet files (pd.read_csv, pd.read_parquet) and includes a FastAPI implementation that accepts external dictionary data via POST requests.
Boundary markers: There are no specific instructions or delimiters provided to the agent to treat processed data as untrusted or to ignore instructions embedded within the datasets.
Capability inventory: The agent has the capability to generate and execute Python scripts, train machine learning models, and perform file system read/write operations.
Sanitization: No explicit data validation or sanitization routines are implemented in the provided code snippets for external data inputs.
[DYNAMIC_EXECUTION]: The skill references model serialization and deserialization, which involves executing code stored in binary formats.
Evidence: In references/mlops_production.md, the skill demonstrates the use of joblib.load('model.pkl') for loading serialized models in a real-time API context. While the skill correctly recommends joblib over pickle, both libraries can execute arbitrary code during deserialization if the source file is untrusted. This is a standard practice in MLOps but represents a point of dynamic execution.

data-scientist