SkillTotal AI Component Security Scan

Version updated for https://github.com/pezhik/skilltotal to version v0.39.0.

  • This action is used across all versions by ? repositories.

Action Type

This is a Composite action.

Go to the GitHub Marketplace to find the latest changes.

Action Summary

SkillTotal is a static analysis tool that identifies potential security risks in AI-related components and projects. It scans files locally without running them on the machine, providing evidence-based findings such as file locations and code snippets. SkillTotal is designed to be 100% offline and safe for untrusted components, offering a determinate and standards-aligned approach to detecting supply-chain risks and capabilities before installation or trust.

What’s Changed

Fixed

  • False negative: published packages were scanned without their code (ruleset 43). dist/ and build/ are skipped for a repository, where they duplicate first-party source checked in beside them. A published npm tarball is the mirror image — .npmignore/files keep the sources out and ship only the build output — so those scans covered nothing but the manifest and the README and reported “no findings”. Measured across the public MCP registry, 11 of 12 sampled npm packages that reported zero findings were in fact shipping shell execution, network egress or filesystem access. Build output is now scanned for package artifacts (npm_package, python_package) and still skipped for git/local sources.

Added

  • Build output establishes capability, never risk. A bundle inlines dependencies, tests and templates into one file, blinding the path-based demotion layers, so a risky_construct or malicious_indicator match found in build output is demoted to needs_review while capability findings are kept. Without this, real packages were rated critical from a security tool’s own credential-path watch list and from a bundled SSRF test assertion.
  • Minified bundles are disclosed, not scanned. A bundled script is one very long line, so a finding in it could not carry the checkable file/line evidence every confirmed finding must have, and its inlined dependencies are not the component’s own code. Such files are now reported as a coverage needs_review entry (“Minified bundle not analyzed”) listing the paths — visible rather than silently absent, and without affecting the score.