The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script scripts/media_optimizer.py uses the eval() function to parse frame rate strings (e.g., "30/1") returned by ffprobe. While this is intended to calculate numeric values from fractional metadata, using eval on any data extracted from file metadata is a sub-optimal practice that could theoretically be exploited if the underlying media utility output is compromised.
[INDIRECT_PROMPT_INJECTION]: The skill provides a significant surface for indirect prompt injection as it is designed to ingest and analyze untrusted external media (images, audio, video, and PDFs). Instructions embedded within these files could potentially be interpreted by the multimodal LLM (Gemini) and passed back to the agent, influencing subsequent actions like shell command execution or file modifications.
Ingestion points: gemini_batch_process.py and document_converter.py take untrusted files via the --files and --input arguments.
Boundary markers: Prompts generally lack strict boundary markers or instructions to ignore embedded commands within the media content.
Capability inventory: The skill has access to Bash, Write, and Edit tools, creating a risk if the LLM output is auto-executed by the agent.
Sanitization: There is no explicit sanitization of text extracted from media before it is returned to the agent context.
[COMMAND_EXECUTION]: The scripts/media_optimizer.py script executes system commands (ffmpeg and ffprobe) using subprocess.run. The implementation uses list-based arguments without shell=True, which is a secure way to handle external processes, but users should be aware that the skill relies on the presence and execution of these binary tools.

ai-multimodal