media-use

Pass

Audited by Gen Agent Trust Hub on Nov 21, 2026

Risk Level: SAFEDATA_EXFILTRATIONEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [DATA_EXFILTRATION]: The skill reads and handles sensitive API keys and session tokens for HeyGen, ElevenLabs, and Google Gemini. These credentials are used to make authenticated requests to their respective service endpoints.
  • [EXTERNAL_DOWNLOADS]: The skill fetches media files from external providers including HeyGen and Openverse. It also downloads large AI models for transcription (Whisper), text-to-speech (Kokoro), and music generation (MusicGen) from HuggingFace and other repositories to a local cache (~/.cache/hyperframes).
  • [COMMAND_EXECUTION]: Multiple scripts spawn subprocesses to run system utilities such as ffmpeg, ffprobe, python, and npx. The skill includes a specialized trusted_program.py utility to ensure that these binaries are executed from trusted system paths and not hijacked from within the project directory.
  • [DYNAMIC_EXECUTION]: The audio engine (bgm.mjs) constructs and executes complex Python scripts at runtime via python -c. This is used to drive local generative models like MusicGen with dynamic parameters.
  • [REMOTE_CODE_EXECUTION]: The skill automatically manages its own runtime environment by performing silent pip install commands for missing dependencies like transformers, torch, and google-genai. While targeting legitimate packages, this represents a runtime remote code execution surface.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted media files, transcripts, and HTML compositions. While it uses structured data and path validation to mitigate risks, the ingestion of external content that influences downstream agent actions represents a potential vulnerability surface.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 21, 2026, 10:39 AM