The Agent Skills Directory

[COMMAND_EXECUTION] (LOW): The skill constructs and executes shell commands (ffmpeg, ffprobe) using user-provided input variables such as $INPUT_FILE and output paths. While the instructions use double quotes (e.g., "$INPUT_FILE"), which mitigates simple whitespace issues, it does not explicitly instruct the agent to sanitize filenames against more sophisticated shell injection attacks (e.g., using backticks, semicolons, or command substitution sequences in filenames).
[DATA_EXPOSURE & EXFILTRATION] (SAFE): The skill operates on local video files and does not include any network operations or access to sensitive system directories (e.g., SSH keys, credentials).
[INDIRECT_PROMPT_INJECTION] (LOW):
Ingestion points: Untrusted data enters the agent context via user-provided file paths and filenames.
Boundary markers: None. The skill relies on standard shell quoting.
Capability inventory: Subprocess execution via ffmpeg and ffprobe, and file writing via cat to /tmp/concat_list.txt.
Sanitization: The skill uses quoted here-docs ('EOF') for the concat list, which prevents variable expansion within the heredoc, a good security practice. However, it lacks explicit sanitization instructions for the filenames themselves.

video-concat