The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes graph data from a variety of external sources, which creates a potential surface for indirect prompt injection through maliciously structured graph attributes or metadata.
Ingestion points: references/io.md describes methods for reading data from multiple file formats (GraphML, JSON, GML, CSV, Shapefiles) and SQL databases.
Boundary markers: The documentation provides guidance on maintaining security boundaries when using SQL queries.
Capability inventory: The skill has extensive graph analysis capabilities and the ability to write result files to the local file system.
Sanitization: The documentation explicitly recommends using parameterized queries when interacting with SQL databases to prevent injection attacks.
[DYNAMIC_EXECUTION]: The documentation references the pickle module for graph serialization, which involves executing code during deserialization.
Evidence: references/io.md provides code examples for using pickle.load() and pickle.dump().
Sanitization: The skill includes a clear security warning stating that pickle should only be used with trusted files because it can execute arbitrary code upon loading.

networkx