Back to skill
Skillv1.2.3
ClawScan security
yolo-vision-tools · ClawHub's context-aware review of the artifact, metadata, and declared behavior.
Scanner verdict
BenignMar 16, 2026, 9:53 AM
- Verdict
- benign
- Confidence
- medium
- Model
- gpt-5-mini
- Summary
- This skill is internally coherent for Ultralytics YOLO workflows (model selection, dataset conversion, environment checks), but it probes the local environment and runs shell commands — so review and run it in an isolated environment if you have sensitive data.
- Guidance
- What to consider before installing/using: - Review code: The repository includes scripts that run shell commands and probe the filesystem (searching for other Python interpreters and cached models). This is expected for an environment checker but read the scripts if you want reassurance. - Run in an isolated environment: Use a disposable VM or dedicated virtualenv/conda environment if you are worried about exposing environment details or interfering with system Python installs. - No secrets requested: The skill does not ask for API keys or credentials. Still, the scripts can reveal local paths, installed package versions, and GPU details — treat that information as potentially sensitive on shared machines. - Network behavior: SKILL.md examples show loading images from URLs and installing ultralytics with pip; when you run the skill, it may download models or fetch remote images if you pass URLs. Only use trusted model/image URLs. - Subprocess/shell usage: The check script uses subprocess.run with shell execution and executes other python interpreters to check for ultralytics. This is normal for a diagnostic tool but increases what the script can inspect. Avoid running it on systems where arbitrary command output must remain private. - Mitigations: Inspect/modify scripts to remove any checks you don't want, run them with restricted permissions, or run only the functions you trust. Prefer running 'pip install -U ultralytics' yourself and invoking well-known commands (yolo checks) rather than running every diagnostic script unmodified. If you want, I can point out the specific lines that probe your home directory, check other python interpreters, or execute shell commands so you can audit them more easily.
Review Dimensions
- Purpose & Capability
- okThe name/description (YOLO vision tools) aligns with the included code and docs: model selection, dataset conversion, training helpers, and environment checks. The files and functions are directly relevant to the stated computer-vision tasks.
- Instruction Scope
- noteSKILL.md instructs installing ultralytics and running environment checks; the provided check_environment.py enumerates Python environments, executes external python interpreters, runs shell commands (via subprocess), and inspects ~/.cache/ultralytics for models. Those actions are reasonable for diagnosing YOLO setups but do extend beyond pure inference (they probe filesystem and other Python installs).
- Install Mechanism
- okThere is no install spec; this is instruction- and script-based (no remote downloads baked into an installer). That reduces supply-chain risk. The SKILL.md suggests installing ultralytics via pip, which is expected for this purpose.
- Credentials
- noteThe skill does not request credentials or environment variables. However, scripts read user paths (home directory caches, virtualenv locations), execute local Python interpreters, and run system commands (which may reveal environment details). These accesses are justifiable for environment checks but are broader than a minimal inference-only tool.
- Persistence & Privilege
- okThe skill does not request persistent privileges, does not set always: true, and does not declare writing to other skills' config. It will save outputs to workspace/yolo-vision or user-specified project folders (normal behavior).