The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process data from various untrusted external sources, including BIM (IFC) files, IoT sensor feeds, and WebSocket messages. These inputs are used to update the properties and status of digital twin elements. The implementation lacks explicit sanitization or boundary markers to differentiate data from instructions, creating a surface for indirect prompt injection where malicious content in a data file might influence the agent's behavior. \n
Ingestion points: SKILL.md (through import_from_ifc, process_sensor_update, and WebSocket message handling). \n
Boundary markers: Absent in the provided Python implementation. \n
Capability inventory: The skill has filesystem permissions (via claw.json) and performs network operations using the websockets library. \n
Sanitization: No data sanitization or validation logic is present for property values or sensor inputs.\n- [EXTERNAL_DOWNLOADS]: The skill's implementation in SKILL.md requires external Python packages numpy and websockets, which are not part of the standard library. These are well-known and widely used packages.\n- [DYNAMIC_EXECUTION]: The DigitalTwinCore class implements a sensor binding mechanism that accepts a transform callable to modify data values at runtime. While intended for unit conversion or scaling, if the agent dynamically generates or assigns these functions based on unvalidated user input, it introduces a risk of executing unintended logic.

digital-twin-sync