visual-style

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

• Third-party content exposure detected (high risk: 0.85). The required runtime path is “Extract” mode: it ingests user-supplied URLs/PDFs/videos, fetches/parses their contents (e.g., from-website/from-pdf/from-video), and then generates style_prompt_full that is fed into the agent LLM context—this is outsider-authored free text from public web content or uploaded documents.

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

• Potentially malicious external URL detected (high risk: 1.00). The "Extract from Website" workflow fetches a user-supplied website URL at runtime (e.g., the example https://stripe.com) and uses the fetched HTML/CSS content to synthesize the style_prompt_full which is then used verbatim by connectors (like HeyGen), so external content directly controls the agent's prompts.