nano-banana-image-ad

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (medium risk: 0.30). In SKILL.md Phase 2–4 and scripts/generate_image.py main()/build_prompt()/generate(), outsider-authored free text (the user’s seed prompt and pin-block description) is rewritten and then sent as the prompt field to Novoads at runtime via POST /v1/estimates and POST /v1/images.

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (high risk: 1.00). The skill explicitly directs runtime fetching of shared prompt and safety files from https://raw.githubusercontent.com/novoads/claude-code-ads/main/ (or by running git clone https://github.com/novoads/claude-code-ads.git), and those fetched files (prompt-library.md, safety-suffixes.md, guide.md) directly control prompt rewriting and safety instructions used by the generator.