media-use

Pass

Audited by Gen Agent Trust Hub on Nov 19, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDATA_EXFILTRATIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill fetches media assets and machine learning models from several external sources:
  • Downloads music and video assets from HeyGen's API (api.heygen.com) and CDN (static.heygen.ai).
  • Fetches sound effects from the Openverse API (api.openverse.org).
  • Accesses the Google Gemini API (generativelanguage.googleapis.com) for text-to-speech generation.
  • Downloads machine learning models (e.g., Parakeet, Whisper, FLUX) from HuggingFace and other official repositories during setup or first run.
  • [COMMAND_EXECUTION]: The skill frequently executes system-level tools to process media:
  • Uses ffmpeg and ffprobe for decoding audio, measuring durations, and transcoding media files.
  • Spawns Python processes to run local generative models (MusicGen) and utility scripts.
  • Implements a robust trusted_program helper that prevents the execution of binaries planted in the project directory by enforcing absolute paths from trusted system folders.
  • [DATA_EXFILTRATION]: The skill transmits usage telemetry to PostHog (us.i.posthog.com). This is documented as anonymous by default and can be opted out of via HYPERFRAMES_NO_TELEMETRY=1. It also handles API keys for HeyGen, ElevenLabs, and Google Gemini, which are sent to the respective service providers to facilitate media generation.
  • [DYNAMIC_EXECUTION]: The skill dynamically generates Python code at runtime to handle MusicGen audio generation. It uses JSON.stringify for safe interpolation and explicitly removes the current working directory from sys.path in the spawned Python process to prevent local module poisoning.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied text for text-to-speech and prompts for image/video generation. While this represents an attack surface for indirect injection, the risk is mitigated by using structured data formats (JSON) and passing content directly to provider APIs rather than interpreting it as instructions for the local agent.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 19, 2026, 08:23 PM