doubao-multimodal

Fail

Audited by Gen Agent Trust Hub on Sep 22, 2026

Risk Level: HIGHCOMMAND_EXECUTIONPRIVILEGE_ESCALATIONEXTERNAL_DOWNLOADSDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill makes extensive use of Bun.spawn to execute shell commands for media processing and network operations.
  • Evidence found in scripts/lib/ffmpeg.ts where ffmpeg and ffprobe are invoked to process audio and video files.
  • Evidence found in scripts/lib/tos.ts where curl is executed to perform HTTP PUT requests for file uploads to Volcano TOS.
  • [PRIVILEGE_ESCALATION]: The skill contains logic to automatically install system dependencies using administrative privileges.
  • Evidence found in scripts/lib/ffmpeg.ts in the ensureFfmpegInstalled function, which executes sudo apt-get, sudo dnf, or sudo pacman to install ffmpeg on Linux systems.
  • [EXTERNAL_DOWNLOADS]: The skill downloads media content from arbitrary remote URLs provided via user input.
  • Evidence found in scripts/lib/source.ts in the downloadRemote function which uses fetch to retrieve files from user-supplied URLs and stores them in a local cache directory.
  • [DYNAMIC_EXECUTION]: The skill dynamically assembles shell commands using variables derived from environment configuration and user-provided file paths.
  • Evidence found in scripts/lib/ffmpeg.ts and scripts/lib/tos.ts where command arrays for Bun.spawn are constructed at runtime using variables like mediaPath, outputPath, and putUrl.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external media (audio/video) and passes derived content to an LLM, creating a potential vector for indirect injection.
  • Ingestion points: Untrusted media files are ingested from local paths (--path) or remote URLs (--url) in scripts/main.ts.
  • Boundary markers: The skill includes a buildPartHeader in scripts/lib/prompts.ts which adds a "Note: this is a segment" message for split media, but it does not provide explicit isolation or "ignore instructions" delimiters for the media content itself.
  • Capability inventory: The skill possesses capabilities for subprocess execution (ffmpeg, curl), local file system writes (Bun.write in utils.ts), and network operations (fetch in ark.ts).
  • Sanitization: There is no sanitization of the audio or video content, as it is processed as binary data before being sent to the LLM service.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Sep 22, 2026, 10:49 AM