video-fetch-and-summarize

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

• Third-party content exposure detected (high risk: 0.90). The skill (SKILL.md and scripts/fetch_and_summarize.py) explicitly downloads videos from arbitrary public URLs (YouTube, Instagram, TikTok, Twitter/X, Reddit, and 1000+ platforms via yt-dlp) and sends the user-generated video content to the Gemini model for summarization, so untrusted third-party content is ingested and can influence generated outputs and downstream decisions.

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

• Potentially malicious external URL detected (high risk: 0.90). The skill extracts and downloads arbitrary user-provided HTTP(S) video URLs (regex: https?://... via extract_urls_from_text) at runtime and then uploads the fetched video file into the Gemini model call (client.models.generate_content(contents=[uploaded, prompt])), thereby injecting external content directly into the model context and potentially allowing remote content to control the generated instructions/output.