The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests data from Blender's compositor node tree, creating a surface for indirect prompt injection if the scene data is maliciously crafted.
Ingestion points: The tools get_compositor_node_value, list_compositor_node_links, and list_compositor_node_types (scripts/get_compositor_node_value.py, scripts/list_compositor_node_links.py, scripts/list_compositor_node_types.py) read socket values, link structures, and node metadata from the active Blender scene.
Boundary markers: There are no explicit boundary markers or instructions to the agent to ignore natural language instructions that might be embedded within node names, socket labels, or string-based socket values.
Capability inventory: The skill possesses significant mutation capabilities within the Blender environment, including creating nodes, deleting nodes, and modifying connections (scripts/create_compositor_node.py, scripts/delete_compositor_node.py, scripts/connect_compositor_nodes.py).
Sanitization: The provided Python scripts do not implement sanitization or validation logic to filter out potentially malicious instruction content retrieved from the Blender scene before it is returned to the agent context.

blender-compositor