byted-mediakit-video

Pass

Audited by Gen Agent Trust Hub on Nov 5, 2026

Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill is designed to interact with the Volcano Engine's official mediakit-cli tool. It executes shell commands to perform legitimate video processing operations, such as metadata probing, frame extraction, and video enhancement. This behavior is documented and forms the core functionality of the skill.
  • [INDIRECT_PROMPT_INJECTION]: The skill's primary purpose involves processing external media from URLs and interpreting user prompts for video understanding (e.g., in video-understand-router.md). This creates an inherent surface for indirect prompt injection where instructions could be embedded in the media content or the user's natural language requests.
  • Ingestion points: video_url, audio_url, and target_face_url parameters in all tools, as well as the prompt parameter in the video understanding tool.
  • Boundary markers: Instructions in SKILL.md mandate that the agent must request missing parameters from the user and avoid forging or inferring values, ensuring user intent is maintained.
  • Capability inventory: The skill has the shell permission and executes the mediakit-cli binary to process ingested data.
  • Sanitization: Inputs are handled as discrete CLI flags, and structured JSON parameters follow specific schemas defined in the reference documentation to prevent command injection.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 5, 2026, 12:59 AM