skills/datadrivenconstruction/ddc_skills_for_ai_agents_in_construction/llm-document-extraction/Gen Agent Trust Hub
llm-document-extraction
Pass
Audited by Gen Agent Trust Hub on Sep 18, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill extracts text and image data from external construction documents (RFIs, submittals, contracts) and includes this content directly in prompts sent to the LLM without sanitization or boundary markers.\n
- Ingestion points: Data enters the system via
extract_from_pdf,extract_from_drawing, andcreate_document_indexinSKILL.md, which read contents from user-provided PDF and image files.\n - Boundary markers: The prompt templates in
SKILL.md(e.g., inextract_from_pdfandquery_document) do not use unique delimiters or explicit instructions to the AI to ignore instructions found within the document text.\n - Capability inventory: The skill utilizes filesystem access to read documents and network access to interact with the OpenAI API and Qdrant vector database.\n
- Sanitization: No sanitization, escaping, or filtering of the document content is performed prior to interpolation into the LLM prompts.
Audit Metadata