AI security review — superpowers

Plugin: obra/superpowers · 282k stars · MIT Reviewer: Fable 5.1 · Date: 6 September 2026 Grade: A · Risk: none

What it is

A core skills library for Claude Code by Jesse Vincent — TDD, debugging, collaboration patterns, and a set of reusable workflows. It ships as a plugin: a plugin.json manifest, fourteen skills, three instruction files, and one lifecycle hook.

What I looked at

I read the manifest and every bundled file — not just the plugin metadata, because a plugin’s real reach is the code and hooks it installs, not the JSON that lists them. That is eighteen files in total.

The manifest is ordinary: name, version, MIT license, author. It declares no MCP servers and runs nothing at install time.

Findings

One item is worth a sentence rather than an alarm. The plugin registers a SessionStart hook that runs its own run-hook.cmd when a session starts. That is the skill loader doing its job, and the command stays inside the plugin’s own directory (${CLAUDE_PLUGIN_ROOT}). It executes on its own, so you should know it is there, but it does not reach outside the plugin or touch anything of yours.

The scanner also flagged rm -rf <workspace> inside the subagent-driven-development skill. That line deletes the scratch workspace the skill itself created for a plan; the git history is kept. It removes its own temporary directory, nothing else.

Nothing else tripped a rule, and a full-text pass for the usual trouble — piped installers, eval, credential reads, exfiltration to a third party — came back clean.

Verdict

A well-known, well-kept engineering skills library. No malware, no data collection, no code that runs against you. The one hook is standard plugin behavior and stays in its own yard. Grade A.

_Produced by agentmods.dev — automated AI security review. Reviewer model: Fable 5.1. Date: 6 September 2026. How this review works: https://agentmods.dev/about#security _