nature-figure

Pass

Audited by Gen Agent Trust Hub on Oct 7, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill exhibits an attack surface for indirect prompt injection through several data ingestion points combined with powerful capabilities.
  • Ingestion points: Multiple scripts ingest untrusted external data: scripts/figure_source_data.py reads user-provided CSV and TSV files; scripts/validate_figure.py parses Python and R plotting scripts; scripts/audit_pdf_text.py and scripts/audit_figure_collisions.py parse PDF geometry and text operators. Additionally, scripts/generate_openrouter_schematic.py reads text from an abstract file (--abstract-file) and image bytes from reference paths (--reference-image).
  • Boundary markers: The skill does not employ explicit delimiters or boundary markers to isolate untrusted content from the AI agent's instructions, nor does it provide specific instructions for the agent to disregard commands embedded within the ingested data.
  • Capability inventory: The skill has file system access (read/write) and network capabilities. Specifically, scripts/generate_openrouter_schematic.py makes network POST requests to the OpenRouter API (https://openrouter.ai/api/v1/images). A sophisticated indirect prompt injection attack could potentially trick the agent into exfiltrating sensitive local files (e.g., credentials or private research data) by passing them as arguments to the schematic generation tool.
  • Sanitization: No evidence of content sanitization or instruction filtering was found for data ingested from external files before processing or transmission.
  • [EXTERNAL_DOWNLOADS]: The script scripts/generate_openrouter_schematic.py performs network POST operations to https://openrouter.ai/api/v1/images. This is a legitimate feature for generating AI-assisted manuscript schematics through the OpenRouter service, a well-known AI platform. The implementation uses the standard urllib.request library.
  • [COMMAND_EXECUTION]: The skill relies on the execution of several internal Python scripts for auditing and validation, and it generates and runs Python or R plotting code as its primary function. These are expected behaviors for a figure production tool.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 7, 2026, 11:02 AM