The Agent Skills Directory

[PROMPT_INJECTION]: The instructions in SKILL.md explicitly direct the agent to include the --dangerously-skip-permissions flag when invoking the agy vision tool, which is a deliberate bypass of built-in security checks.
[COMMAND_EXECUTION]: The skill executes shell commands via the subprocess module in scripts/_vision.py and scripts/anibon/ytdlp.py to call yt-dlp and ffmpeg. These calls process external URLs, potentially leading to command injection.
[INDIRECT_PROMPT_INJECTION]: The skill ingests automated transcripts from external videos to trigger analysis. The lack of sanitization for this untrusted data creates a surface where instructions embedded in video content could influence the agent.
[DYNAMIC_EXECUTION]: The library file scripts/anibon/markdown.py uses __import__ for dynamic module loading, and SKILL.md provides complex code snippets for the agent to execute at runtime for media extraction.

antigravity-vision-proxy