math-figure-generator
Pass
Audited by Gen Agent Trust Hub on Sep 3, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from local data sources (CSV, JSON, MAT files) and modeler-provided claims (from
qx_figure_table_plan.md). This creates a surface where embedded instructions in data could influence agent behavior, though the skill implements safeguards. - Ingestion points: Figure claims are pulled from
methods/Qx/qx_figure_table_plan.mdor decision logs; numeric data is loaded from external files via user-defined paths. - Boundary markers: The skill mandates the use of sentinel markers like
[AI-DRAFT]and[MODELER INPUT NEEDED]to identify unconfirmed claims and prevent their promotion. - Capability inventory: The skill utilizes file writing capabilities to save figures (
save_figure) and append to local logs (render_check_and_log). - Sanitization: Instructions explicitly forbid the AI from originating judgment-bearing claims and require pulling from human-confirmed sources.
Audit Metadata