AI Agent Audit — Security skill for Claude Code

ai-agent-audit

Read-only security audit of a local developer machine, focused on the attack surface that AI coding agents can reach.

This tool answers a single question: **"What can an AI coding agent see and do on this machine right now, and how bad would it be if one of them went off the rails?"**

It scans for credentials, tool configuration, hooks, MCP servers, agent skills, IDE extensions, session artifacts, and supply-chain compromise indicators that AI agents (Cursor, Claude Code, Gemini CLI, Cline, Aider, Codex, Windsurf, Amazon Q, and others) routinely have access to. Every finding maps to a documented incident class.

TL;DR

Read-only scan of `$HOME` for everything an AI coding agent can reach: SSH keys, cloud creds, npm/pip tokens, MCP server configs, AI hooks, repo-level `.claude/.cursor/` configs, agent skills, IDE extensions, plus credential leakage in shell history and session JSONL/SQLite. Outputs a security score (0–100) and five reports — Markdown, JSON, classified secrets inventory with redacted fingerprints, prioritised action plan, and a cyberpunk-themed HTML report that auto-opens in your browser. No network, no auto-remediation, no raw secret values written anywhere.

Quick start

**Install the skill globally** (works from any project, every Claude Code session):

git clone https://github.com/bartek-filipiuk/ai-agent-audit.git ~/.claude/skills/ai-agent-audit
chmod +x ~/.claude/skills/ai-agent-audit/scripts/run-audit.sh \
         ~/.claude/skills/ai-agent-audit/scripts/aggregate.sh \
         ~/.claude/skills/ai-agent-audit/scripts/lib/*.sh \
         ~/.claude/skills/ai-agent-audit/scripts/modules/*.sh

The `SKILL.md` at the repo root is what Claude Code auto-discovers.

**Run directly without an agent:**

bash ~/.claude/skills/ai-agent-audit/scripts/run-audit.sh
# HTML report auto-opens in your browser at the end.

**Or — let Claude Code handle it.** Drop one of these prompts into any session:

Run the ai-