doubao-multimodal

Pass

Audited by Gen Agent Trust Hub on May 7, 2026

Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONDATA_EXFILTRATION
Full Analysis
  • [COMMAND_EXECUTION]: The skill uses Bun.spawn to execute several system commands, including ffmpeg and ffprobe for media transcoding, probing, and splitting, as well as curl for uploading files to Volcano TOS.
  • [EXTERNAL_DOWNLOADS]: The downloadRemote function in scripts/lib/source.ts allows the skill to download media files from arbitrary URLs provided by the user via the --url flag.
  • [REMOTE_CODE_EXECUTION]: The ensureFfmpegInstalled function in scripts/lib/ffmpeg.ts provides a routine to install the ffmpeg package using system package managers (brew, apt-get, dnf, pacman). While this involves executing installation scripts, it is presented as a prerequisite setup step and generally requires user confirmation unless the --yes flag is explicitly provided.
  • [DATA_EXFILTRATION]: The skill is designed to read local files (provided via --path) and upload them to a Volcano TOS bucket. While this is the intended primary function of the tool for processing local media with the Ark API, it represents a capability that handles sensitive data (user media) and transmits it to a remote cloud service based on user-provided credentials.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data such as subtitles and transcripts which are interpolated into prompts for the Ark multimodal model.
  • Ingestion points: Untrusted data enters the context via --subtitle, --transcript, --prompt, and media file content (audio/video).
  • Boundary markers: The skill uses delimiters like 「<transcript>」 in scripts/lib/prompts.ts to separate user-provided text from system instructions.
  • Capability inventory: The skill has command execution capabilities via Bun.spawn for ffmpeg, ffprobe, and curl (used for media processing and TOS uploads).
  • Sanitization: Shell arguments for external commands are sanitized using quoteShellArg or passed as argument arrays to Bun.spawn, which mitigates shell injection risks.
Audit Metadata
Risk Level
SAFE
Analyzed
May 7, 2026, 07:05 AM
Security Audit — agent-trust-hub — doubao-multimodal