figure-spec

Warn

Audited by Gen Agent Trust Hub on Nov 10, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [DYNAMIC_EXECUTION]: The skill calculates the path to the executable renderer script (figure_renderer.py) at runtime by parsing a local file (.aris/installed-skills-codex.txt) and searching through multiple environment-dependent paths (~/.codex/, $ARIS_REPO). This dynamic path resolution can be subverted if an attacker modifies the configuration files.
  • [INDIRECT_PROMPT_INJECTION]: The skill facilitates an automated review loop where user-influenced data (the FigureSpec JSON and SVG) is passed to a secondary AI agent (spawn_agent). This creates an attack surface where malicious content embedded in the diagram labels or metadata could influence the behavior of the reviewer agent.
  • Ingestion points: Diagram descriptions are taken from $ARGUMENTS and external files like PAPER_PLAN.md to generate the JSON spec.
  • Boundary markers: No explicit boundary markers or 'ignore embedded instructions' warnings are included when passing the generated files to the reviewer agent.
  • Capability inventory: The reviewer agent is granted high capabilities (e.g., gpt-6-astra with xhigh reasoning effort) and access to the rendered file paths.
  • Sanitization: While the renderer validates the JSON schema, it does not sanitize the semantic content of labels or metadata against instructional injections.
  • [COMMAND_EXECUTION]: The skill explicitly uses the Bash tool to execute python3 for rendering and validation, and rsvg-convert for PDF generation. These operations rely on the integrity of the dynamically resolved $FIGURE_RENDERER variable.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Nov 10, 2026, 09:36 AM