Skillv1.0.0

ClawScan security

Whisper Transcriber · ClawHub's context-aware review of the artifact, metadata, and declared behavior.

Scanner verdict

BenignMar 8, 2026, 10:05 AM

Verdict: Benign
Confidence: high
Model: gpt-5-mini
Summary: This package is internally coherent: it is an offline Whisper (whisper.cpp) wrapper that expects ffmpeg and whisper-cli, installs packages / downloads model files, and the requested actions line up with the stated purpose.
Guidance: This skill appears to do what it says, but review and be cautious before running the installer and scripts: 1) The installer will call your system package manager (may prompt for sudo) and will download large model files from https://huggingface.co/ggerganov/whisper.cpp — ensure you have sufficient disk space and trust that source. 2) config.json contains empty modelsSha256 entries, so by default the provided integrity check has no expected hashes; the scripts will only verify if those fields are filled and the host has shasum/sha256sum (the small node -e helper used for verification requires node to be present). 3) Prefer running install.sh and transcribe.sh manually (inspect them first), do not run them as root unless you understand the package installs, and consider running inside an isolated environment (VM or container) if you want extra safety. 4) If you plan to enable autonomous agent invocation, remember the skill will download models and run local binaries on any audio it processes — ensure your agent's permissions and inputs are appropriate. If you want, I can point out the exact lines in the scripts that perform downloads, package installs, and sha256 checks so you can inspect them before running.

Review Dimensions

Purpose & Capability: okName/description require ffmpeg and whisper-cli and perform audio conversion + offline ASR; the scripts, download behavior, and config match that purpose. Required binaries and model downloads are appropriate for an offline whisper.cpp-based transcriber.
Instruction Scope: noteSKILL.md and scripts instruct the agent/user to run install.sh and transcribe.sh which install packages, download models, create temp dirs, convert audio, and invoke whisper-cli. The instructions are scoped to the transcribe task, but the runtime scripts will: (1) call system package managers (may use sudo), (2) download multi-hundred-MB/GB model files from huggingface, (3) optionally use node -e to read config.json for sha256 verification. None of these expand scope beyond transcription, but they do perform privileged installs and large external downloads — review before running.
Install Mechanism: okThere is no platform install spec in the registry; the repo provides cross-platform install.sh that uses package managers (brew/apt/dnf/yum/pacman/zypper) and curl/wget to download models from huggingface.co (a well-known host). No URL shorteners or personal IPs observed. This is a common, expected install mechanism for model-based tools.
Credentials: okThe skill declares no required credentials and only optional environment variables for model/temporary directories and defaults (WHISPER_DEFAULT_MODEL, WHISPER_MODEL_DIR, WHISPER_TEMP_DIR, etc.). Scripts do not attempt to read unrelated secrets or system credentials.
Persistence & Privilege: okalways:false and user-invocable. The skill does not request permanent platform-level presence or modify other skills' config. Install scripts create files under the skill's assets/models and may write system packages via package managers (normal for installers).