The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill manages authentication for the HeyGen service by reading secrets from standard local paths such as ~/.heygen/credentials and .env files. These credentials are used exclusively for authorized API calls to the service and are not exfiltrated to unauthorized domains. A public PostHog ingestion key is included in the telemetry module, which is a standard practice for this service and does not pose a security risk.
[EXTERNAL_DOWNLOADS]: To support BGM generation and local Fallbacks, the skill automatically installs Python packages from PyPI (e.g., transformers, torch, google-genai, elevenlabs) when dependencies are missing. These installations target well-known libraries required for the skill's primary functionality.
[COMMAND_EXECUTION]: The skill makes extensive use of external CLI tools, including ffmpeg, ffprobe, heygen, mflux-generate, codex, parakeet-mlx, and ltx-2-mlx. All process spawns use argument arrays rather than shell strings, effectively mitigating shell injection risks. A dedicated regression test suite ensures that filenames with shell metacharacters cannot trigger unintended command execution.
[DATA_EXFILTRATION]: Usage telemetry is sent to PostHog (us.i.posthog.com) to track media resolution success and provider utilization. The data is anonymized, excludes PII like intent text or local file paths, and honors standard opt-out environment variables such as DO_NOT_TRACK and HYPERFRAMES_NO_TELEMETRY.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied natural language descriptions to drive generative AI tools (TTS, Image Gen). It provides a surface for prompt injection; however, it mitigates risk by treating these inputs as data for specialized CLI tools and APIs rather than executing them as code, and includes SSRF protections to prevent the agent from being directed to internal or malicious network resources.

media-use