The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The setup documentation in references/setup.md explicitly suggests that the user "Request an unsandboxed run" if platform security restrictions (sandboxing) prevent the browser from launching or accessing GPU resources. Encouraging the removal of security boundaries is a significant risk.
[COMMAND_EXECUTION]: The skill relies heavily on executing external binaries for media processing. The scripts media.mjs, render.mjs, mix.mjs, takes.mjs, and pick.mjs use execFileSync, spawn, and spawnSync to run ffmpeg, ffprobe, python, and whisper. The arguments for these commands are derived from storyboard.json and lines.json, which could be manipulated if sourced from untrusted inputs.
[DYNAMIC_EXECUTION]: In scripts/takes.mjs, a multi-line Python script is dynamically constructed as a string and executed using spawnSync with the -c flag. This pattern of runtime code generation and execution is a common vector for injecting arbitrary logic.
[EXTERNAL_DOWNLOADS]: The skill requires downloading and installing multiple external packages and binaries during the setup phase, including Playwright, FFmpeg, and Python libraries for speech and transcription. Additionally, the chatterbox-tts library is documented to download model weights from Hugging Face on its first run.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-controlled text from lines.json and storyboard.json, which is then rendered using innerHTML within a Playwright-controlled browser instance in assets/template/engine.js. This creates a vulnerability where malicious content in the input files could lead to cross-site scripting (XSS) or other execution within the rendering browser environment.

marketing-video