The Agent Skills Directory

[DYNAMIC_EXECUTION]: Multiple scripts, including add_node.py, create_scene.py, gltf_processor.py, load_sprite.py, and ui_assembler.py, generate temporary GDScript files (.gd) at runtime. These files are then executed using the Godot engine's headless mode via the GodotBase.run() method defined in scripts/base.py.
[COMMAND_EXECUTION]: The generated scripts use Python f-strings to interpolate user-provided command-line arguments (such as args.type, args.name, args.res, and args.json) directly into the executable code without sanitization. This facilitates classic code injection. For example, in scripts/add_node.py, the variable args.type is used to instantiate a node: var n = {args.type}.new(). An attacker could provide a payload that breaks out of this statement to execute arbitrary GDScript logic, which has full access to the host operating system through Godot's OS.execute() and file system APIs.
[INDIRECT_PROMPT_INJECTION]: The skill processes project assets and external data (JSON layouts, scene resources). If an attacker provides a maliciously crafted Godot project or data file, the skill will ingest this content and incorporate it into the vulnerable code-generation process. This allows for unauthorized code execution when the agent performs automation tasks on a compromised project file.

godot-builder