The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill manages user credentials for HeyGen and Google Cloud services, resolving secrets from environment variables or local configuration files located in ~/.heygen/ and ~/.hyperframes/. Secrets are used to authenticate requests to media generation APIs.
[EXTERNAL_DOWNLOADS]: Fetches color grade LUTs from HeyGen's CDN (static.heygen.ai), retrieves sound effects from the Openverse API, and downloads synthesized media from provider-specific signed URLs. A dedicated fetch utility implements a blocklist to prevent SSRF against internal networks.
[REMOTE_CODE_EXECUTION]: The skill's runtime environment facilitates the installation of third-party Python packages (transformers, torch, google-genai, etc.) from PyPI using pip to enable local generative capabilities for background music and transcription.
[COMMAND_EXECUTION]: Spawns system processes for media processing (ffmpeg, ffprobe), language interpretation (python3), and Node.js package execution (npx). Subprocess arguments are constructed from user-supplied intents, though generally handled as data arrays rather than raw shell strings.
[DYNAMIC_EXECUTION]: Employs dynamic script generation for media tasks, such as background music creation using MusicGen. Template-generated Python code is executed via the python3 -c flag, with user inputs escaped using JSON.stringify to prevent code injection.
[DATA_EXFILTRATION]: Transmits anonymous usage telemetry (event names, media types, and provider selection) to a PostHog instance. The implementation explicitly excludes sensitive metadata like intent text or file paths and filters out user IP addresses.
[INDIRECT_PROMPT_INJECTION]:
Ingestion points: User-supplied intents, audio script text, and JSON request files enter the context from untrusted sources.
Boundary markers: Uses structured JSON for data exchange and applies standard escaping before content interpolation into command arguments or scripts.
Capability inventory: Capability to execute shell commands and Python interpreters with inputs derived from ingested data.
Sanitization: Inputs are primarily sanitized via JSON serialization and the use of argument arrays in spawn and execFileSync calls to avoid shell interpretation.

media-use