The Agent Skills Directory

[COMMAND_EXECUTION]: The skill extensively uses child_process methods (spawn, execFileSync) to invoke external utilities such as ffmpeg, ffprobe, npx, and the heygen CLI. The implementation avoids shell injection vulnerabilities by passing arguments as arrays rather than concatenated strings.
[REMOTE_CODE_EXECUTION]: To support its audio and image generation features, the skill automatically installs required Python packages (transformers, torch, google-genai, etc.) using pip if they are not detected on the system. While these are reputable libraries, the auto-installation from remote registries represents a runtime code execution vector.
[EXTERNAL_DOWNLOADS]: Media assets, color LUTs, and AI models are downloaded from external sources, primarily HeyGen's CDN and Google's Developer APIs. A dedicated utility validates these URLs against a network blocklist to prevent SSRF attacks against internal infrastructure.
[DATA_EXFILTRATION]: The skill transmits anonymous usage telemetry (media types, providers used, and success rates) to PostHog for analytics. While it Link usage to user accounts if signed in to HeyGen, it explicitly excludes sensitive content like intent text or local file paths from these reports.
[INDIRECT_PROMPT_INJECTION]: The skill ingests and processes user-supplied storyboard and brief files. This data is interpolated into command-line arguments and prompt templates for downstream AI models, creating a potential surface for indirect injection attacks, although the skill includes several sanitization routines.

media-use