vibe-hardening

Version updated for https://github.com/vibe-hardening/cli to version v0.4.0.

• This action is used across all versions by 0 repositories.

Action Type

This is a Composite action.

Go to the GitHub Marketplace to find the latest changes.

Action Summary

Summary:
The vibe-hardening GitHub Action is a security scanner designed for AI-generated codebases, particularly those built with popular AI coding tools. It automates the detection of vulnerabilities such as hardcoded secrets, injection risks, insecure authentication, and supply chain issues across multiple languages (JavaScript, TypeScript, Python, Go, Rust). By providing rule-based analysis and tailored outputs, it helps developers identify and mitigate security risks efficiently, ensuring safer and more reliable code.

What’s Changed

vibe-hardening 0.4.0 — new agent scan subcommand for AI agent skill files.

npx vibe-hardening agent scan

Why now

Speed is the only moat in indie security tooling. The first OpenClaw / Hermes / Cursor skill compromise that hits HN front page will trigger Google searches for “agent skill scanner” — and the tool that’s already shipped, ranked, and starred wins that traffic. Built before the first skill compromise hits the news.

What it does

Statically scans local AI agent skill files for security issues. Auto-detects 10 platforms by filesystem stat:

Plus generic ~/.<agent>/skills/ pattern for the long tail.

5 rule packs / 65 detection rules

• A — Hardcoded secrets in SKILL.md / configs / .env. Reuses v1’s 27 SECRET_RULES verbatim. Hermes specifically stores all secrets in ~/.hermes/.env (not config.yaml) — that file is now scanned.
• B — Prompt injection patterns (11 rules): “ignore previous instructions”, role overrides, ChatML control tokens, Llama instruction tags, zero-width hidden characters.
• C — Dangerous shell commands (14 rules) in body and scripts/ files: rm -rf /, curl | sh, eval/exec on user variables, persistence into .bashrc / authorized_keys, fork bombs.
• D — Skill schema / body integrity (5 active sub-rules): missing required fields, hidden scripts/ directory, sensitive path + nearby exfil verb, env-dump pattern, folder name typosquatting popular skills.
• G — MCP server config issues (6 rules): http:// (non-TLS), localhost residue, secrets in env block, server name typosquatting (Levenshtein ≤ 2), > 20 servers, npx -y of unverified packages.

CLI flags

vh agent scan                            # all detected platforms
vh agent scan --target cursor            # filter to one platform
vh agent scan --rule b,c                 # only rule packs B + C
vh agent scan --exclude secret           # skip secret rules
vh agent scan --severity high            # CI-friendly threshold
vh agent scan --format json -o r.json    # machine-readable
vh agent scan --no-telemetry             # one-shot opt-out

Telemetry

Sends a separate event_type: 'agent_scan' event with an agents_detected presence vector across known platforms. Reuses the same opt-in config as v1 (default off, never re-prompts). Zero PII.

Stats

• 65 new agent-scan rules
• 74 existing code-scan rules
• 406 tests
• 192 KB tarball, 268 files
• MIT, free forever

Install

npx vibe-hardening agent scan         # one-shot
npm install -g vibe-hardening         # global

Full changelog: CHANGELOG.md.

• Vibe-Hardening
• GitHub Actions