The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The script scripts/transcribe.py (lines 39, 51) automatically executes pip install at runtime if the rich or tqdm libraries are missing. This initiates unauthorized downloads and modifies the user's Python environment without a confirmation prompt at the point of failure.
[EXTERNAL_DOWNLOADS]: The installation script scripts/install-requirements.sh downloads and installs multiple Python packages from PyPI, including faster-whisper, openai-whisper, tqdm, and rich. It also attempts to install system-level dependencies like ffmpeg via Homebrew on macOS.
[DATA_EXFILTRATION]: In scripts/transcribe.py (lines 302, 310) and SKILL.md (lines 359, 367), the full text of generated audio transcripts is sent to external command-line interfaces (claude and gh copilot). This represents a significant privacy risk if the audio files contain sensitive or confidential information, as the content is transmitted to external AI providers for summarization.
[COMMAND_EXECUTION]: The skill relies heavily on the subprocess module to execute shell commands. This includes running pip for package management, ffmpeg/ffprobe for audio processing, and calling external LLM CLI tools. While arguments appear to be handled as lists, the execution of these external binaries is a core part of the skill's capability.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection. It ingests untrusted data from audio transcriptions and interpolates it directly into prompts sent to external LLMs (e.g., scripts/transcribe.py line 291) without sufficient boundary markers or sanitization. A malicious actor could embed instructions within an audio file that are then executed by the downstream LLM.
[SAFE]: Mentions of API keys and credentials in references/tools-comparison.md (e.g., AZURE_SPEECH_KEY, ASSEMBLYAI_API_KEY) are clearly documented as placeholders and examples for optional third-party service integration and do not contain hardcoded secrets.

audio-transcriber