The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to execute shell commands using user-supplied project names and filenames as arguments. Specifically, it uses commands like mkdir -p "projects/<active-project-name>/clips" and mv ~/Downloads/<original-filename>.mp4 .... This pattern is vulnerable to command injection if the input contains shell metacharacters (e.g., ;, &&, |).\n- [DATA_EXFILTRATION]: The skill is designed to access and move files from the user's ~/Downloads/ directory. While intended for importing video clips, this involves automated access to a directory that often contains sensitive personal or professional information.\n- [EXTERNAL_DOWNLOADS]: The skill directs users to download and install a desktop application from an external GitHub repository (github.com/storytold/artcraft).\n- [REMOTE_CODE_EXECUTION]: The skill executes a local bash script ./scripts/new-project.sh and passes user-provided arguments directly to it.\n- [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection due to the lack of sanitization on ingested data. Ingestion points: User-provided intentions, project names, and reference files (read via the Read tool). Boundary markers: No explicit delimiters or instructions are used to separate untrusted data. Capability inventory: Access to Bash (mkdir, mv, cp, open, ls), Write (files), and Read (images). Sanitization: No validation of user-provided strings before use in commands or file paths.

ai-film-director