The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script scripts/media_optimizer.py uses the eval() function to calculate the frame rate (fps) from the r_frame_rate field provided by ffprobe. Because this metadata is derived from the media file being optimized, a maliciously crafted file containing Python code in its frame rate metadata could lead to arbitrary code execution when the optimization utility is run.
[COMMAND_EXECUTION]: The skill executes external binaries (ffmpeg and ffprobe) using subprocess.run within scripts/media_optimizer.py. Although arguments are passed as a list to mitigate shell injection, the invocation of system-level utilities on untrusted files represents a significant capability that could be exploited if vulnerabilities exist in the underlying media libraries.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection attack surface.
Ingestion points: Multimedia files (audio, video, images) and PDFs processed through scripts/gemini_batch_process.py and scripts/document_converter.py.
Boundary markers: The prompts sent to the LLM do not utilize explicit delimiters or instructions to ignore instructions embedded within the source media content.
Capability inventory: The skill is explicitly granted Bash and Write tool access in SKILL.md.
Sanitization: Extracted text and transcriptions from untrusted documents and media are not sanitized before being interpolated into LLM prompts.
[EXTERNAL_DOWNLOADS]: The skill references and installs official, well-known libraries such as google-genai (Google's official Python SDK), Pillow, and pypdf. These are considered safe sources.
[SAFE]: API key management follows industry best practices, instructing the user to store the GEMINI_API_KEY in environment variables or local .env files that are prioritized correctly by the scripts.

ai-multimodal