The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill documents and promotes the use of Python's pickle module and NetworkX's read_gpickle function for graph serialization.
Evidence: Found in references/io.md under the 'Pickle Format' and 'Format Selection Guidelines' sections.
Risk: The pickle module is insecure. If the agent is instructed to load a pickle file provided by an untrusted external source, it could lead to remote code execution.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process graph data from a wide variety of external file formats and databases, creating a surface for indirect prompt injection.
Ingestion points: Includes nx.read_edgelist, nx.read_graphml, nx.read_gml, nx.read_adjlist, and pd.read_sql_query documented in references/io.md.
Boundary markers: The instructions do not define clear boundaries or 'ignore' instructions for the data being processed from these files.
Capability inventory: The skill environment allows for complex Python execution, file system access (read/write), and integration with numerical and visualization libraries.
Sanitization: There is no evidence of sanitization or validation of the graph attributes or node/edge data loaded from external sources before they are processed by the agent.

networkx