d-data-analysis-semantic-layer
Pass
Audited by Gen Agent Trust Hub on Oct 12, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill executes a local Python script
scripts/validate_candidate.pyto validate candidate records before they are committed to the semantic layer. This script uses only standard libraries and performs schema and logic validation. - [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process 'candidate conclusions' provided by users or other tasks. This represents a potential indirect prompt injection surface. However, the skill mitigates this through several layers of security:
- Ingestion Points: Ingests external candidate data files (JSON/JSONL) via the
scripts/validate_candidate.pyscript. - Boundary Markers: The instructions require the user to explicitly see the report and approve the content using specific verbal triggers ('Confirm entry into knowledge base', etc.) before any writing occurs.
- Capability Inventory: Capabilities are limited to reading local reference files and performing file writes using host-provided handles; no arbitrary network operations or shell access are present outside the fixed validation script.
- Sanitization: The validation script enforces strict data types, allowed values for status and fact types, and performs logical checks on dates (e.g., expiry must be after review).
Audit Metadata