The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from the system clipboard and local save files to instantiate game objects, creating a potential attack surface if malicious data is provided. \n
Ingestion points: scripts/sandbox_world_serializer.gd (local JSON file at user://sandbox_world.json) and scripts/blueprint_manager.gd (system clipboard).\n
Boundary markers: None identified in the provided scripts.\n
Capability inventory: Uses load(), instantiate(), and add_child() to create nodes from data; uses FileAccess for local disk operations.\n
Sanitization: No validation or path sanitization is performed before loading resource paths from the data source.\n- [DYNAMIC_EXECUTION]: The skill dynamically loads and instantiates Godot scenes based on paths stored in external data sources.\n
Evidence: scripts/sandbox_world_serializer.gd uses load(entry.res) where entry.res is a string sourced from a parsed JSON file.\n
Context: While this is a standard and necessary implementation for save-game systems and blueprint sharing in the sandbox genre, it theoretically allows for the loading of arbitrary scenes if the data source is compromised.

godot-genre-sandbox