Six posts into this series and I’ve mentioned installing a lot of third-party skills and plugins without much scrutiny beyond a GitHub star count. Prism Scanner exists precisely because that’s a bad habit – it’s an open-source security scanner built specifically for auditing agent skills, plugins, and MCP servers before they get anywhere near your machine.
What it checks for
Rather than a generic static analysis pass, it runs over 39 rules aimed at the specific ways a malicious or careless skill can behave badly: prompt injection vectors buried in a SKILL.md, scripts that reach out to unexpected network endpoints, overly broad tool permissions, and similar agent-specific risks that a normal code linter wouldn’t catch. It uses AST-based taint tracking rather than pattern matching alone, and produces an A–F grade for whatever it’s pointed at.
Why this matters for a series like this one
Every skill covered in this series so far – humanizer, Superpowers, detect-ai/humanize, Align, the Karpathy CLAUDE.md – is a third party’s code and instructions running inside an agent that has shell access, file access, and often network access. A scanner built for that specific threat model is a sensible thing to run before installing any of them, this one included.
Where to find it
Prism Scanner turns up as part of the broader claude-code-plugins-plus-skills marketplace, which is worth a browse in its own right if you want to see what an audited, catalogued skill ecosystem looks like at scale – it currently lists over 400 plugins and thousands of individual skills.
Next Sunday: Pluginpool, a small grab bag of sharp, single-purpose developer skills – commit-narrator, pr-storyteller, deps-doctor, and a few others in the same spirit as humanizer.
Leave a Reply
You must be logged in to post a comment.