The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run to call FFmpeg for video processing and FFprobe for metadata extraction. These operations are core to the skill's stated purpose.
[EXTERNAL_DOWNLOADS]: The skill configures container images that install software from trusted sources, including Debian's apt repositories (FFmpeg), NVIDIA's registry (CUDA), and the Python Package Index (Torch, Transformers, Accelerate, etc.).
[DATA_EXFILTRATION]: One production pattern in references/production-audio-troubleshooting.md includes functionality to send job status metadata to an external URL via an HTTP POST request using curl. While intended for job monitoring, this network operation targets user-defined external endpoints.
[PROMPT_INJECTION]: The skill processes external media files (video/audio) which serves as an ingestion point for potential indirect prompt injection attacks against the underlying processing tools.
Ingestion points: transcode_video in SKILL.md and transcribe_video in references/production-audio-troubleshooting.md ingest raw bytes from external sources.
Boundary markers: No explicit delimiters or boundary markers are used to isolate untrusted media data.
Capability inventory: The skill possesses subprocess execution (FFmpeg, FFprobe, Curl) and persistent file system access (Modal Volumes).
Sanitization: The skill uses list-based arguments for subprocess.run which prevents shell injection, but it does not sanitize or validate the internal contents of processed media files.

ffmpeg-modal-containers