The Agent Skills Directory

[DYNAMIC_EXECUTION]: In scripts/long-to-short.sh, the $out variable is directly interpolated into a JavaScript string executed by node -e. This creates a code injection vulnerability where a maliciously crafted filename or path containing a single quote could execute arbitrary JavaScript code.
[DYNAMIC_EXECUTION]: The scripts/anim.sh script executes complex logic using the node -e command. While arguments are passed via process.argv rather than direct interpolation, the use of dynamic execution for primary logic remains a significant attack surface.
[COMMAND_EXECUTION]: Scripts such as ken-burns.sh, long-to-short.sh, and stamp-cta.sh contain a resolve_cli function that searches for the capcut CLI in parent node_modules directories and the system PATH. This searching behavior can be abused to execute a malicious binary if it is placed in one of the searched paths.
[DATA_EXFILTRATION]: As described in api-reference.md, the capcut caption command includes a --cloud option that uploads audio data to ByteDance's servers for transcription. This constitutes an external data transmission to the vendor's cloud infrastructure.
[EXTERNAL_DOWNLOADS]: The skill documents capabilities for fetching assets from Wikimedia Commons (e.g., commons.wikimedia.org) via the add-video and add-audio commands. The implementation includes license classification and restricts network activity to Wikimedia-specific domains.

capcut-edit