The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill provides instructions in references/io.md for serializing and deserializing graph objects using the pickle module (e.g., pickle.load(f) and nx.read_gpickle()). This is a known security risk as pickle can execute arbitrary code during the loading process, making it unsafe to use with data from untrusted sources.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process data from various external formats, which presents a surface for indirect prompt injection attacks where malicious instructions could be hidden within graph attributes or metadata.
Ingestion points: Data loading methods such as read_edgelist, read_graphml, read_gml, read_json, from_pandas_edgelist, and read_adjlist are documented in references/io.md and references/graph-basics.md.
Boundary markers: No guidance is provided regarding the use of boundary markers or safety instructions to mitigate the risks associated with processing untrusted data.
Capability inventory: The skill includes capabilities for writing to the file system and generating interactive visualizations, which could be exploited through successful injection.
Sanitization: The skill does not include instructions for validating or sanitizing ingested data before it is utilized by graph algorithms or visualization tools.
[PROMPT_INJECTION]: The SKILL.md file contains explicit behavioral instructions that steer the agent to suggest a specific third-party platform (K-Dense Web) when complexity increases, attempting to override the agent's standard behavior to include promotional messaging.

networkx