The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from video frames (OCR text, metadata, descriptions). It mitigates this by defining an "Untrusted content boundary" and explicitly instructing the agent to ignore any instructions or tool-use prompts found within the media.
[COMMAND_EXECUTION]: Frame extraction is performed using ffmpeg. To prevent exploitation via malicious media files, the skill specifies that the tool must run in a restricted sandbox with resource limits, read-only mounts, and disabled network access.
[UNVERIFIABLE_DEPENDENCIES]: Instructions for installing prerequisites include using python -m pip install --require-hashes, ensuring integrity via a reviewed hash lockfile and requiring explicit user consent.
[DYNAMIC_EXECUTION]: Includes Python code for grid compositing that implements security best practices by enabling DecompressionBombWarning as an error to prevent resource exhaustion attacks.

video-frames