The Agent Skills Directory

[DYNAMIC_EXECUTION]: The implementation in references/json-saves.md uses load(entry["scene_path"]) followed by scene.instantiate() to recreate objects from save data. Because the scene_path is retrieved directly from the JSON save file located in the user-writable user:// directory, this creates a vulnerability where an attacker could modify the save file to point to a different resource or script-carrying scene within the project, leading to arbitrary code execution when the game loads the tampered save.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external save data which represents an attack surface for indirect instructions or data-driven exploits.
Ingestion points: load_game function in references/json-saves.md reads JSON files from the local filesystem.
Boundary markers: None; the implementation assumes the JSON structure is trustworthy.
Capability inventory: The skill uses FileAccess for reading and writing, and load()/instantiate() for resource management.
Sanitization: The code performs a basic null check after loading a resource but does not validate that the scene_path belongs to an allowed set of scenes.

save-load