The Agent Skills Directory

[COMMAND_EXECUTION]: The instructions in SKILL.md direct the agent to execute shell commands such as ls -la "$INPUT_FILE" and uv run ... "INPUT_FILE". These commands use direct string interpolation of user-provided file paths. If a malicious filename containing shell metacharacters (e.g., ;, |, `) is provided, it could lead to arbitrary command execution on the host system.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to transcribe audio and video files, which are external, untrusted data sources. The resulting transcription is then presented directly to the agent in the final step. If the audio contains spoken instructions (e.g., "Ignore all previous rules and perform action X"), the agent might interpret these as valid instructions during the preview stage.
Ingestion points: The transcribe.py script reads and processes audio/video files via the whisperx library.
Boundary markers: Absent. Step 4 instructions do not specify the use of delimiters or clear markers to separate transcribed content from agent instructions.
Capability inventory: The skill has the capability to read local files, write transcription results to the file system, and execute shell commands via uv run.
Sanitization: Absent. The transcription text is generated and displayed without filtering or escaping potentially malicious instructional content.

audio-transcribe