bim-cost-estimation-cwicr

Pass

Audited by Gen Agent Trust Hub on Sep 17, 2026

Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [COMMAND_EXECUTION]: The skill uses subprocess.run to call a local executable at C:\DDC\RvtExporter.exe. This is used to convert Revit/IFC models into quantity take-off (QTO) data. While restricted to a specific vendor path, executing local binaries is a high-privilege action.
  • [INDIRECT_PROMPT_INJECTION]: The skill has a vulnerability surface for indirect prompt injection where data from external BIM models can influence the AI's behavior.
  • Ingestion points: BIM element names and categories are read from Excel files (exported from models) in SKILL.md via pd.read_excel.
  • Boundary markers: Absent. Data is directly embedded into the decompose_bim_type prompt using f-strings without delimiters or warnings to the model.
  • Capability inventory: The skill can execute local commands via subprocess.run, write files to the local system using pd.ExcelWriter, and perform network operations to OpenAI and Qdrant APIs (all in SKILL.md).
  • Sanitization: Absent. No escaping or validation is performed on the bim_type or category strings before they are sent to the LLM.
  • [EXTERNAL_DOWNLOADS]: The skill depends on several external Python packages: pandas, numpy, qdrant-client, openai, and openpyxl.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 17, 2026, 08:19 AM