The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads media assets and color grading files from HeyGen's domains (heygen.ai and static.heygen.ai). It also retrieves brand logos and icons from public registries and repositories including svgl.app, cdn.jsdelivr.net, github.com, and icons.duckduckgo.com. AI model weights are fetched from HuggingFace repositories.
[COMMAND_EXECUTION]: The skill makes extensive use of local binaries for media processing and AI inference, including ffmpeg, ffprobe, python, npx, and specific model runners like ltx-2-mlx and mflux-generate. The implementation uses direct binary execution with tokenized argument arrays for operations involving user-supplied data (such as search intents), which effectively prevents shell injection attacks.
[DATA_EXFILTRATION]: Operational telemetry is sent to a PostHog instance (us.i.posthog.com). The data collected is limited to metadata such as media type, provider, and resolution source. When a user is signed into the HeyGen service, the skill uses their account email or username as a distinct identifier. This behavior is documented, and opt-out environment variables (HYPERFRAMES_NO_TELEMETRY, DO_NOT_TRACK) are supported.
[INDIRECT_PROMPT_INJECTION]: The skill processes natural language descriptions of media needs which are passed to various providers. While this constitutes an attack surface, the inputs are used as parameters for specific search or generation tools rather than being interpreted as system-level instructions, mitigating the risk of behavioral overrides.
[CREDENTIALS_UNSAFE]: The skill manages API keys and OAuth tokens for HeyGen, ElevenLabs, and Google services. These secrets are retrieved from standard locations such as environment variables and the user's local configuration directory (~/.heygen), which is standard practice for tools integrating with these services.

hyperframes-media-use