model-sample-image-export
Pass
Audited by Gen Agent Trust Hub on Jul 14, 2026
Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface by ingesting model checkpoints and benchmark results from the results/ directory to update documentation.
- Ingestion points: The agent reads trained checkpoints and benchmark run artifacts from the results/ folder (SKILL.md).
- Boundary markers: The skill lacks technical delimiters (e.g., XML tags) for external data, relying on natural language instructions to verify and inspect quality.
- Capability inventory: The agent has the authority to write image files to docs/source/images/ and modify README.md and other documentation files.
- Sanitization: No automated sanitization is specified; the process relies on the agent's internal evaluation of the output quality.
- [COMMAND_EXECUTION]: The skill requires the agent to generate and execute custom scripts to export data and may involve running model training tasks.
- Evidence: The skill instructs the agent to "derive a narrow exporter" for specific models and "schedule a few runs" if existing checkpoints are insufficient.
Audit Metadata