The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill implements runtime input rebinding persistence using potentially unsafe deserialization methods.
scripts/safe_runtime_rebind.gd uses bytes_to_var() to reconstruct InputEvent objects from a file stored in user://input_rebinds.cfg.
scripts/input_remapper.gd uses str_to_var() to deserialize event data from user://input_bindings.cfg.
While common for local settings in Godot, these functions can be leveraged for unsafe object construction if the underlying configuration file is modified by a malicious actor on the local system.
[INDIRECT_PROMPT_INJECTION]: The input rebinding system creates an attack surface by ingesting external data from the local file system.
Ingestion points: Files located at user://input_rebinds.cfg and user://input_bindings.cfg are read at runtime in load_rebinds() and load_bindings().
Boundary markers: None identified; the data is directly passed to deserialization functions.
Capability inventory: The skill has the capability to modify the engine's InputMap globally using InputMap.action_add_event(), which can fundamentally alter how the agent or the application responds to user actions.
Sanitization: The skill performs basic type checks (e.g., typeof(encoded) != TYPE_PACKED_BYTE_ARRAY or event is InputEvent) but does not fully validate the integrity or the specific properties of the deserialized objects before applying them to the InputMap.

godot-input-handling