The Agent Skills Directory

[PROMPT_INJECTION]: The skill exhibits a vulnerability to indirect prompt injection due to its handling of untrusted external inputs.
Ingestion points: Untrusted data enters the agent context through the file_path, text, and ops parameters in tool.py.
Boundary markers: The skill does not implement delimiters or specific instructions to the model to ignore potential commands embedded within audio transcripts or processed text.
Capability inventory: The skill has the capability to execute system commands via FFmpeg, perform file system operations (read/write), and communicate with external APIs for speech services.
Sanitization: While path validation exists in validate_file_path, it uses a blacklist of system directories (/etc/, /proc/, etc.) which may leave other sensitive user directories unprotected.
[COMMAND_EXECUTION]: The skill executes external FFmpeg processes using subprocess.run. Although it uses argument lists to prevent shell-based injection, the dynamic construction of command flags from user-provided parameters like ops or model represents an attack surface for argument injection.
[EXTERNAL_DOWNLOADS]: The skill automates the installation of FFmpeg and multiple Python dependencies such as openai-whisper, gTTS, and librosa. These downloads originate from well-known technology organizations and standard package registries.

audio-processing