The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill possesses an attack surface where untrusted data in the form of JSON specifications can influence agent behavior or Blender state.
Ingestion points: Data enters through the --spec parameter in scripts/gn_builder.py, which reads and parses JSON files from the filesystem.
Boundary markers: No boundary markers or validation logic exists to differentiate between intended data and potentially malicious instructions within the JSON specification.
Capability inventory: The skill can create node trees, modify object attributes, and save .blend files to the local file system using the _bpyx.py helper script.
Sanitization: There is no sanitization of the nodes or links fields in the JSON; strings are passed directly to Blender's nodes.new() and setattr() functions, which could be exploited to instantiate unexpected node types or modify internal object properties.
[COMMAND_EXECUTION]: The skill instructs the agent to run Blender in background mode to execute Python scripts with arbitrary arguments. While this is the intended mechanism for the skill, it provides a powerful interface for file system and application state manipulation that could be abused if parameters are influenced by malicious input.

blender-geometry-nodes