The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill instructs the agent and scripts to use yt-dlp --cookies-from-browser chrome. This command extracts session cookies directly from the user's local browser database (Chrome) to authenticate with YouTube, which constitutes a high-risk exposure of sensitive credentials and session tokens.
[OBFUSCATION]: The SKILL.md file contains the control character U+0008 (Backspace) embedded within markdown code block headers (e.g., ash instead of bash). This character is invisible in many editors and is used to obfuscate the shell environment being invoked.
[PRIVILEGE_ESCALATION]: In scripts/extract_visual_livechat.py, the agy vision tool is invoked with the --dangerously-skip-permissions flag. This flag is designed to explicitly bypass security prompts or permission checks, allowing the script to perform restricted operations without user oversight.
[INDIRECT_PROMPT_INJECTION]: The skill defines a workflow where raw, untrusted YouTube chat messages are processed and analyzed by subagents to determine 'Thai Subculture & Psychology'. There are no boundary markers or sanitization procedures implemented, making the analysis subagents vulnerable to instructions hidden within the viewer chat replay data.
[COMMAND_EXECUTION]: The script scripts/extract_visual_livechat.py makes extensive use of subprocess.run to execute external binaries including ffmpeg, ffprobe, yt-dlp, and agy. The commands are constructed using dynamic input from the range argument, which, although partially validated, represents a command injection attack surface.

anibon-livechat-analysis