thumbnail

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

• Third-party content exposure detected (medium risk: 0.30). The skill ingests outsider-authored free text via Step 1/30-36 where user-supplied {keyword} and competitor video descriptions are used to query the Supabase competitor_videos table, and Step 41-48 then surfaces untrusted thumbnail_url items derived from that data for subsequent downloading/analysis.

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

• Potentially malicious external URL detected (high risk: 0.90). The skill downloads a competitor thumbnail at runtime via the curl command (curl -sL "{thumbnail_url}" ...) — e.g. the example YouTube thumbnail URL https://i.ytimg.com/vi/.../maxresdefault.jpg — and the fetched image is then analyzed to build the generation prompt, so external content directly controls the agent's prompts.