Bastion Skill — Security scanner for AI agent skills

Scan third-party AI agent skills for malicious code before you install them.

Problem

Agent skills, tools and MCP plugins increasingly come from strangers. Users already recognize that scanning an AI agent skill for malicious code is a necessary step, but there is no standard audit path, so many either hand-read the code or install blindly.

Solution

An open-source CLI plus GitHub Action that statically scans a skill directory for known red flags: outbound domains and IPs, sensitive file access, shell execution, background daemons, env var exfiltration and obfuscation. It outputs a readable report with severity levels and accepts custom YAML rule packs. No sandbox execution to avoid over-engineering.

Users

Developers and platform teams who install many third-party agent skills locally or in production.

MVP

A CLI that parses skill/plugin layouts and flags suspicious network and file access; ships with about 20 built-in danger rules; prints a color terminal report plus JSON; drops into GitHub Actions as a merge gate; rule packs open to the community as YAML.

Revenue

Free open-source rule engine; enterprise subscription adds team rule sync, CI integration, monthly threat reports and priority rule updates.

Why now

Agent skill and MCP marketplaces are exploding but auditing habits lag behind. Pure static analysis means light dependencies, and one person can ship a usable first version in one to two weeks.