The Agent Skills Directory

[Dynamic Execution] (MEDIUM): The file references/io.md includes code examples for using pickle.load() and nx.read_gpickle(), which are vulnerable to arbitrary code execution.
Evidence: Code snippets such as G = pickle.load(f) and nx.read_gpickle(G, 'graph.gpickle') demonstrate loading serialized Python objects from the filesystem.
Risk: The pickle module is inherently insecure. If an agent follows these examples to process a file provided by an untrusted source, it could lead to full system compromise through remote code execution (RCE).
[Indirect Prompt Injection] (LOW): The skill documents a large surface area for ingesting untrusted data from various file formats and databases.
Ingestion points: references/io.md describes methods to read from Adjacency lists, Edge lists, GML, GraphML, JSON, CSV, and SQL databases.
Boundary markers: The provided examples do not demonstrate the use of boundary markers or instructions to ignore embedded commands within the graph data.
Capability inventory: The skill allows for significant filesystem interaction (reading/writing) and database operations via sqlite3.
Sanitization: There is no evidence of data sanitization or schema validation in the provided documentation examples.

networkx