The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The setup.sh script executes sudo apt-get install to install the ffmpeg package. This requires administrative privileges to modify the system environment during the setup process.
[INDIRECT_PROMPT_INJECTION]: The skill processes narrator scripts and QA questions that are interpolated into model prompts without sanitization, creating a surface for indirect injection.
Ingestion points: Narrator lines from video/vo/narrator.json and user-provided questions for the listen.mjs script.
Boundary markers: The prompt in tts.mjs uses a simple "Read:\n" prefix but lacks robust delimiters or instructions to the model to ignore potential commands embedded in the text.
Capability inventory: The skill can execute shell commands via ffmpeg and node, perform network operations via fetch and WebSockets, and control a browser via agent-browser.
Sanitization: Input text is not validated or escaped before being included in model system instructions or user queries.
[EXTERNAL_DOWNLOADS]: setup.sh downloads font assets from Google's official GitHub repository (github.com/google/fonts) to provide typography for video captions.
[DATA_EXFILTRATION]: Scripts such as tts.mjs, listen.mjs, and publish.mjs read the ~/.openkbs/projectJWT file to retrieve authentication tokens for use with the vendor's API and proxy services.
[DYNAMIC_EXECUTION]: compose.mjs uses the dynamic import() statement to load the scene configuration from video/scenes.mjs, which allows for the execution of arbitrary JavaScript code provided in the configuration file.
[COMMAND_EXECUTION]: The skill makes extensive use of shell command execution, including calling ffmpeg, ffprobe, node, and agent-browser to perform video processing and recording tasks.

demo-video