The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process UI layouts and actions defined in JSON format received from remote servers or local files. This creates an attack surface where a malicious backend could provide instructions aimed at influencing the agent or the application's behavior.
Ingestion points: Layout data is ingested via XDriver.remote and XDriver.static as documented in SKILL.md and references/public_api.md.
Boundary markers: The instructions do not define specific delimiters or "ignore embedded instructions" warnings for the processed JSON data.
Capability inventory: The framework utilizes network operations via HttpTransportManager, native module interaction via NativeModuleCapabilityDelegate, and scripting via ScriptingCapabilityDelegate (as noted in references/capabilities.md).
Sanitization: No explicit sanitization or validation logic for the data content is provided in the instructions.
[DYNAMIC_EXECUTION]: The skill documentation mentions the ScriptingCapabilityDelegate in references/capabilities.md, which is intended for "Embedded script execution". This feature allows the framework to execute logic provided within the layout data, which constitutes dynamic execution at runtime.

flutter-duit-bdui