DEV Community

Yurukusa
Yurukusa

Posted on

I Ran a Safety Scan on My Claude Code Setup — Here's What I Found

#claudecode #ai #safety #devtools

I've been running Claude Code autonomously for 200+ hours. Built hooks, wrote CLAUDE.md, set up a multi-agent system where two AIs consult each other while I sleep.

I assumed my setup was solid.

Then I wrote a 10-item diagnostic script and ran it on my own environment. Two items were red.

The 10 Things That Can Go Wrong

These aren't theoretical. Each one came from a real failure during autonomous operation:

  1. No CLAUDE.md — The AI has no persistent instructions. Every session starts from scratch.
  2. No hooks installed — No automated checks run after tool calls. Errors go unnoticed.
  3. No dangerous command protection — rm -rf, git reset --hard execute without warning. Replit's AI deleted a production database in 2025.
  4. No git auto-backup — No automatic branch creation before risky changes. One bad edit and you're recovering from reflog.
  5. No session state saving — When context runs out, all progress disappears. No checkpoint, no handoff notes.
  6. No external action gate — The AI can push code, post comments, send messages without approval.
  7. No error tracking — Errors happen, get fixed with a band-aid, and the root cause is never recorded.
  8. No secrets in .gitignore — .env, credentials, API keys sitting in the repo, one git push from public.
  9. No settings.json — Claude Code's permission system isn't configured. Default behavior may not match your expectations.
  10. Risky operations in git history — git reset --hard or git clean -fd in your reflog means it happened at least once.

Each item has a weight (1-3 points). Total possible score: 19.

My Results

I ran the scan on my main project directory:

Risk Score: 2/19 (MODERATE)
Enter fullscreen mode Exit fullscreen mode

Two items were red: .gitignore didn't have secret patterns, and git auto-backup wasn't hooked up. Everything else passed — because I'd already spent months building the hooks.

But here's the thing: I built those hooks after the accidents happened. The dangerous command hook exists because Claude Code ran rm -rf on a project directory. The context monitor exists because a session died at 3% context with no checkpoint.

What a Fresh Setup Looks Like

On a clean machine with no hooks, no CLAUDE.md, no configuration:

Risk Score: 16/19 (CRITICAL)
Enter fullscreen mode Exit fullscreen mode

9 out of 10 items fail. The only thing that passes is "no risky operations in git reflog" — because nothing has happened yet.

What Changes After the Fix

The diagnostic has a --fix flag that installs 4 free safety hooks:

curl -sL https://gist.githubusercontent.com/yurukusa/10c76edee0072e2f08500dd43da30bc3/raw/risk-score.sh | bash -s -- --fix
Enter fullscreen mode Exit fullscreen mode

After running it:

Risk Score: 7/19 (HIGH)

Improved by 9 points.
  Before: 16/19 (CRITICAL)
  After:  7/19 (HIGH)
Enter fullscreen mode Exit fullscreen mode

The free hooks cover: CLAUDE.md creation, hook installation, dangerous command detection, and settings.json setup. The remaining items (git backup, session saving, external gates, error tracking, .gitignore, reflog history) need additional configuration.

Try It

Run the scan (read-only, nothing installed):

curl -sL https://gist.githubusercontent.com/yurukusa/10c76edee0072e2f08500dd43da30bc3/raw/risk-score.sh | bash
Enter fullscreen mode Exit fullscreen mode

10 seconds. Runs locally. No data sent anywhere. Source code on GitHub.

If you want to fix what it finds:

curl -sL https://gist.githubusercontent.com/yurukusa/10c76edee0072e2f08500dd43da30bc3/raw/risk-score.sh | bash -s -- --fix
Enter fullscreen mode Exit fullscreen mode

Existing files are never overwritten.

More Resources

The free hooks handle the basics. If you're running extended autonomous sessions with multiple agents, the CC-Codex Ops Kit adds multi-agent relay, stall detection, watchdog processes, and a task queue system — the infrastructure for sessions that run while you're away.

Top comments (0)