The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill provides implementation examples for using the pickle module to load graph data in references/io.md. Deserializing data with pickle.load() is inherently unsafe as it can execute arbitrary code contained within the file. While the documentation includes a warning to only unpickle files from trusted sources, the inclusion of this pattern without prioritizing secure alternatives represents a risk of arbitrary command execution.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest data from numerous external file formats, which can act as a vector for instructions embedded in data attributes.
Ingestion points: Graph data is read through several interfaces in references/io.md, including nx.read_graphml, nx.read_gml, pd.read_sql_query, and various text-based loaders.
Boundary markers: The instructions do not define boundary markers to isolate data content from the agent's core instructions, nor do they advise the agent to disregard instructions found within graph properties.
Capability inventory: The skill includes functions that can write files to disk (nx.write_graphml, pd.to_csv), perform database operations (pd.to_sql), and generate visualizations (plt.savefig).
Sanitization: There is no mention of sanitizing node or edge attributes before processing, which could allow malicious content to persist through the analysis pipeline, although the documentation correctly recommends parameterized queries for SQL safety.

networkx