skill-trust-auditor

To perform a security audit before installation, fetch the skill's SKILL.md and referenced scripts, then scan them using deterministic regex pattern matching to identify malicious shell, network exfiltration, and prompt injection indicators.

A Trust Score is a 0–100 rating generated by evaluating a skill's scanned content against known attack vectors, converting regex pattern matches into a verdict like SAFE, INSTALL WITH CAUTION, RISKY, or DO NOT INSTALL.

Yes, you can use an optional LLM-as-judge advisory mode powered by Anthropic Claude Haiku to analyze ambiguous curl intent when an Anthropic API key is available in your environment.

The scanner requires Python3 and the requests library to fetch and scan skill content, while the anthropic dependency is only needed if you enable the optional LLM-as-judge advisory mode.

Pre-install checks detect supply chain risks including secret-harvesting, shell injection, network exfiltration, and prompt injection indicators by running 52 regex-based checks across common script types.

Regex pattern matching provides deterministic detection of known malicious patterns but may flag ambiguous network behavior, requiring optional LLM-as-judge advisory analysis to accurately assess complex curl intent.