explainer-video

Warn

Audited by Gen Agent Trust Hub on Nov 12, 2026

Risk Level: MEDIUMDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [DYNAMIC_EXECUTION]: The file scripts/video.mjs uses await import() with a URL constructed dynamically from either an environment variable (ARTICLE2MEDIA_ENGINE) or a relative file path (../../article2media/engine/scripts/). This dynamic loading from computed paths allows the skill to execute code located outside of its own directory, which is a significant security finding as the behavior depends on the state of the local filesystem or environment variables.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes external, untrusted articles and instructs the agent to act as a "director" to interpret the content and generate a video.json script. This workflow is vulnerable to instructions embedded in the source article that could influence the agent's output or tool usage during the scripting phase.
  • Ingestion points: article.md (the input article provided by the user or fetched from a URL).
  • Boundary markers: The instructions lack explicit boundary markers or warnings to the agent to ignore instructions embedded within the source content.
  • Capability inventory: The skill utilizes the Bash tool to execute scripts that perform file system operations, invoke the ffmpeg tool, and interact with external APIs.
  • Sanitization: The skill relies on the agent's natural language processing to summarize the content into a structured JSON format without formal sanitization of the input text.
  • [COMMAND_EXECUTION]: The skill makes extensive use of the Bash tool to run various sub-commands of the video engine, including doctor for environment checks, check for script validation, and make for the full rendering pipeline. These commands are executed based on instructions provided in the SKILL.md and video-guide.md files.
  • [EXTERNAL_DOWNLOADS]: The documentation indicates that the skill may utilize external services such as edge-tts (network-based text-to-speech) and various AI models (OpenAI, Gemini) for illustration generation and voice cloning. While these are documented as part of the skill's functionality, they involve network operations to external endpoints.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Nov 12, 2026, 04:48 PM