AI Skill Safety Checker
Inspect skill instructions and selected text files for patterns that deserve manual review. Nothing is executed or uploaded.
Choose the text to inspect
Up to 100 text files, 2 MB of selected files and 2 million characters per batch. Archives and binaries are not inspected.
No review yet.
Review findings in context
No review yet.
What the rules look for
This original, rule-based checker searches for remote command execution, attempts to bypass permissions, instruction overrides, sensitive paths or credentials, destructive operations, network destinations, dependency installation, persistence and dynamic or encoded code. It reports file names, line numbers and short evidence excerpts.
High review priority means a pattern deserves attention; it is not a maliciousness verdict. A warning such as “never run this command” can match the same rule as an instruction to run it. Read the surrounding source and follow the actual data flow.
How to audit a package
- Obtain a specific version from a source you can identify. Keep its license and provenance.
- Read the manifest and choose its supporting scripts, references and dependency files. The checker cannot discover unselected files or follow remote links.
- Run the local review, inspect every relevant match and compare it with the surrounding code. Mentioned domains are not automatically contacted.
- Trace what the package reads, writes, downloads, executes and sends. Review install hooks and versioned dependencies separately.
- Test an accepted candidate only with disposable data and suitable permissions. Record the exact version and your decision.
Limits of this review
No configured patterns matching does not mean a skill is safe. Obfuscation, indirect imports, compiled code, an omitted file or a changed remote dependency can hide behavior. This is not antivirus, a complete code audit, a vulnerability scan or an AI judgment. It does not execute scripts, unpack ZIP files, retrieve URLs or resolve dependencies.
The review stays in this page memory until you reset or close it; it is not saved to an account. Downloaded reports can contain the excerpts you supplied, so keep them with the reviewed project. Normal site analytics and cookies are covered by the privacy policy; this implementation does not send selected input text or file names to analytics.
A useful next step
Frequently asked questions
Does a report with no findings prove safety?
No. The configured patterns can miss behavior and cannot inspect unselected or remote files.
Are links fetched or scripts executed?
No. The tool reads selected text and reports matches locally.
Can comments cause findings?
Yes. Examples, warnings and negated instructions can match. Interpret every result in context.
Made by AI Craft Pad
Original implementation, version 1.0. Free, no account or API key required. About the publisher · Contact or report an issue.