nano-banana-image-ad
Pass
Audited by Gen Agent Trust Hub on Sep 15, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill utilizes a local Python script (
scripts/generate_image.py) to manage the image generation workflow, including file handling and API communication. - [EXTERNAL_DOWNLOADS]: The skill interacts with the
external-api.arcads.aidomain to upload user-provided reference images via presigned URLs and subsequently download the generated ad creatives. - [INDIRECT_PROMPT_INJECTION]: The skill acts as a proxy for user-supplied prompts and images, passing them to an external generative AI service.
- Ingestion points: User-defined prompts and local image reference paths are processed in
SKILL.mdand passed to the helper script. - Boundary markers: The helper script automatically appends strict formatting and safety suffixes (
NO_CHROME_SUFFIX,SAFE_ZONE_SUFFIX,GLYPH_SAFETY_SUFFIX) to the prompt to control model behavior. - Capability inventory: The skill possesses network communication capabilities (
urllib.request) and local file system write access for saving generated images. - Sanitization: The skill relies on the agent to rewrite prompts according to guidelines and the API provider's internal safety filters, but does not perform separate validation on user input content.
Audit Metadata