project-data
Pass
Audited by Gen Agent Trust Hub on Mar 17, 2026
Risk Level: SAFEPROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION]: The skill exhibits an indirect prompt injection vulnerability surface (Category 8) because it is designed to ingest, process, and retrieve data from potentially untrusted external sources, such as uploaded construction plans, specifications, and AutoCAD DWG files. If these source documents contain adversarial instructions, they could potentially override agent constraints when the processed data is retrieved to answer natural language queries. \n
- Ingestion points: Data is ingested from various construction documents (Plans, Specs, Schedules) via extraction pipelines described in references/data-flow-map.md. \n
- Boundary markers: The query guides in references/natural-language-query-guide.md map user questions to data fields but do not specify the use of delimiters or escaping when presenting retrieved content back to the agent context. \n
- Capability inventory: The skill has extensive read/write access to a 28-file JSON data store and is integrated with other skills capable of script and binary execution. \n
- Sanitization: references/extraction-validation-checklist.md outlines rigorous integrity and type validation (dates, IDs, numeric ranges) but does not define sanitization logic to detect or filter malicious natural language instructions within the data.
Audit Metadata