The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection as it ingests user-supplied prompts and reference images for processing.
Ingestion points: User prompts and reference image paths are collected in SKILL.md (Phase 2) and passed as arguments (--prompt, --image-ref, --source) to the execution script.
Boundary markers: The skill incorporates hardcoded safety suffixes (NO_CHROME_SUFFIX, SAFE_ZONE_SUFFIX, GLYPH_SAFETY_SUFFIX) which are appended to prompts in scripts/generate_image.py to constrain the model's output behavior.
Capability inventory: The skill has network access (via urllib.request), file read capabilities (for credentials and reference images), and file write capabilities (outputting to ./generated/). No subprocess executions were found in the functional logic.
Sanitization: The slugify function in scripts/generate_image.py sanitizes prompt text before using it in local file naming to prevent directory traversal or filename injection.
[EXTERNAL_DOWNLOADS]: The skill connects to an external API to perform its primary function and downloads resulting assets.
Evidence: The script scripts/generate_image.py interacts with external-api.arcads.ai to submit generation jobs, upload reference files via presigned URLs, and download the final image results. This is consistent with the skill's stated purpose.

nano-banana-image-ad