The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill fetches content from external, untrusted URLs to generate video scripts and prompts, creating a vulnerability to indirect prompt injection where malicious web content could influence the agent's behavior.
Ingestion points: Untrusted data enters the agent context through the WebFetch and capture_website tools, which retrieve text and screenshots from the user-supplied url parameter.
Boundary markers: The instructions lack explicit delimiters or clear "ignore embedded instructions" warnings for the fetched data, though they do use a claims_allowlist to try and ground the model in specific facts.
Capability inventory: The skill has access to powerful generative tools including generate_reference_video, edit_text_overlay, and add_captions, which could be manipulated by injected instructions.
Sanitization: The skill implements a robust "Post-flight quality gate" using analyze_media to perform OCR on the final video and verify that no unauthorized or garbled claims were introduced during the generation process.

ugc-ads