The Agent Skills Directory

Indirect Prompt Injection (HIGH): The skill processes untrusted external video files while possessing command execution capabilities, which can be exploited if the processing tools (kinemotion/FFmpeg) have vulnerabilities or if the video metadata contains injection payloads.
Ingestion points: Video files processed via CLI commands (<video>, videos/*.mp4) and Python API calls in SKILL.md.
Boundary markers: Absent. There are no instructions to the agent to treat video data as untrusted or to ignore embedded metadata.
Capability inventory: Allowed tools include Bash(uv run kinemotion:*), Read, and Glob. This provides a shell execution path for the results of the data processing.
Sanitization: Absent. While ffmpeg is suggested for rotation issues, there is no security-focused sanitization of the input files.
Unverifiable Dependencies (MEDIUM): The skill relies on the kinemotion package, which is not from a trusted source or organization listed in the security policy.
Evidence: Use of uv run kinemotion in analysis commands and the Python API from kinemotion import ... in SKILL.md.
Command Execution (LOW): The skill requires shell access (Bash) to function. While scoped to kinemotion:*, this still allows execution of arbitrary arguments passed to the third-party binary.

video-analysis-workflow