The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill's authentication logic in scripts/lib/heygen.mjs reads sensitive credentials from ~/.heygen/credentials and searches for .env files in parent directories to resolve API keys for services like HeyGen and ElevenLabs.
[DATA_EXFILTRATION]: Credentials retrieved from the local file system are transmitted to external endpoints, such as api.heygen.com, via network requests in scripts/lib/heygen.mjs to authenticate media generation tasks.
[EXTERNAL_DOWNLOADS]: The skill manages its environment by automatically installing required Python libraries (e.g., google-genai, transformers, torch) via pip at runtime in scripts/lib/bgm.mjs if they are not already present.
[DYNAMIC_EXECUTION]: The skill assembles Python code as string literals and executes it using python3 -c for media processing tasks in scripts/lib/bgm.mjs and scripts/lib/tts.mjs.
[COMMAND_EXECUTION]: The skill spawns subprocesses to execute various system tools and CLI utilities, including ffmpeg, ffprobe, python3, and npx hyperframes.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from audio_request.json and external transcript files (SRT, VTT, JSON), which are then processed by the media engine. These ingestion points have access to command execution and network capabilities, and while transcript-handling.md mentions basic cleanup, the skill lacks robust sanitization and boundary markers for all processed data.

hyperframes-media