genviral

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

• Third-party content exposure detected (high risk: 0.85). In Meta Ads Stage 1 (LP Analysis), the skill fetches an outsider-supplied landing page URL at runtime and extracts readable text/metadata (title, meta description, OG tags, and body text) into runs/*/lp-content.md and runs/*/brand.json, which are then used as LLM context for later stages.

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

• Potentially malicious external URL detected (high risk: 0.90). The meta-ads campaign pipeline fetches arbitrary landing pages at runtime (e.g., lp_analyzer.py calls fetch_url on user-supplied URLs such as "https://genviral.io") and the Meta Ad Library API ("https://graph.facebook.com/v21.0/ads_archive" / "https://www.facebook.com/ads/library/async/search_ads/") is queried at runtime — the scraped LP content and ad data are saved (brand.json, lp-content.md, competitor-ads.md) and directly fed into template/prompt builders (stage prompts and generate_creatives.py), so external URLs clearly control the prompts used by the agent.