nano-banana-image-generator

Pass

Audited by Gen Agent Trust Hub on Mar 15, 2026

Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [PROMPT_INJECTION]: The skill exhibits a surface for indirect prompt injection (Category 8).\n
  • Ingestion points: In SKILL.md, the workflow encourages the agent to use WebSearch or WebFetch to find high-resolution reference images from external domains such as Wikipedia or professional press sites.\n
  • Boundary markers: While the skill uses structured prompt templates (e.g., CONCEPT:, STYLE:), it lacks explicit delimiters or instructions to disregard instructions potentially embedded in metadata or surrounding text of retrieved external assets.\n
  • Capability inventory: The accompanying script scripts/generate_image.py can write files to the local filesystem, make network requests to the Gemini API, and execute local system commands via subprocess.\n
  • Sanitization: No sanitization or validation logic is present to filter content retrieved from external web sources before it is incorporated into the agent's context or generation prompts.\n- [COMMAND_EXECUTION]: The scripts/generate_image.py script utilizes subprocess.run to call the macOS open command. This is used to display generated images in the default system application, which is a common productivity feature.\n- [EXTERNAL_DOWNLOADS]: The skill relies on well-known and trusted external resources, including the google-genai and pillow Python libraries, and the official Google Gemini API endpoints.
Audit Metadata
Risk Level
SAFE
Analyzed
Mar 15, 2026, 05:03 AM