The Agent Skills Directory

[DYNAMIC_EXECUTION]: The documentation and code examples in references/io.md and references/graph-basics.md demonstrate the use of pickle.load() and the nx.read_gpickle() function for serializing and deserializing graph data. The Python pickle module is inherently insecure and can be exploited to execute arbitrary code if it processes a maliciously crafted file.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest data from numerous external and potentially untrusted formats, including GraphML, GML, JSON, CSV, and SQL databases. Malicious instructions could be embedded within graph attributes (node or edge data) to influence the agent's behavior during analysis.
Ingestion points: Multiple functions in references/io.md, such as nx.read_graphml, nx.read_gml, nx.node_link_graph, and pd.read_sql_query, facilitate the loading of external data into the agent's active context.
Boundary markers: The instructions do not define clear delimiters or specify that data attributes should be treated as untrusted, increasing the risk of the agent obeying embedded commands.
Capability inventory: The skill allows for file system operations, database interaction, and dynamic execution via deserialization routines.
Sanitization: No sanitization or validation protocols are described for the content of graph attributes before they are processed by the agent.
[COMMAND_EXECUTION]: The skill references integration with external visualization tools like Graphviz via pydot and pygraphviz in references/visualization.md. These integrations involve spawning subprocesses to run external binaries. If filenames or layout parameters are sourced from untrusted data without validation, it could create a surface for command injection attacks.

networkx