Every app gets reviewed before you install it. AI skills should too.

每個 App 上架都要審核。AI Skill 不該例外。

Quick Start | Online Scanner | ATR Standard | Ecosystem Report

AI agents have full system access -- read files, execute commands, access credentials -- with zero review process. We scanned 36,394 MCP skills. Of 9,676 with source code, 1 in 7 triggered CRITICAL or HIGH severity rules. Credential harvesting. Prompt injection. Data exfiltration. Live numbers.

AI Agent 擁有完整系統權限，卻沒有任何審核。我們掃描了 36,394 個 MCP skills，每 7 個就有 1 個觸發 CRITICAL 或 HIGH security rules。即時數據見 panguard.ai。

Quick Start

npm install -g @panguard-ai/panguard && pga up

One command. Auto-detects your AI platforms, scans all installed skills, starts 24/7 monitoring with real-time dashboard.

一行搞定。自動偵測 platform、掃描所有 skills、啟動 24/7 monitoring + real-time dashboard。

16 platforms: Claude Code, Claude Desktop, Cursor, OpenClaw, Codex, WorkBuddy, NemoClaw, ArkClaw, Windsurf, QClaw, Cline, VS Code Copilot, Zed, Gemini CLI, Continue, Roo Code

Method	Command
npm (recommended)	`npm install -g @panguard-ai/panguard && pga up`
curl (no Node required)	`curl -fsSL https://get.panguard.ai \| bash`
Homebrew (macOS)	`brew install panguard-ai/tap/panguard && pga up`

Or scan online at panguard.ai -- paste a GitHub URL, get a report in 3 seconds.

Screenshots

Dashboard — Security Score, Skills, Protection Status

More screenshots

Welcome — First-run initialization

Rules — 71 ATR detection rules, community contribution

Settings — 3-layer detection (ATR, Fingerprint, Cloud AI)

How It Works

  You install PanGuard
       |
       v
  pga up → scans all skills against 71 ATR rules (< 60 seconds)
       |
       v
  Guard starts → watches for new skill installs in real time
       |
       v
  New skill installed → auto-scanned instantly
       |
       v
  Threat found → anonymized hash uploaded to Threat Cloud
       |
       v
  3+ independent confirmations + LLM review → new ATR rule
       |
       v
  Rule pushed to ALL users within 1 hour
       |
       v
  Next time: blocked in < 50ms → everyone is safer

One person encounters a threat. It becomes a rule. It protects everyone. That's the flywheel.

Three-Layer Detection

Layer	Engine	Latency	Cost
1	71 ATR regex rules	< 50ms	$0
2	Local AI (Ollama)	~ 2s	$0
3	Cloud AI (Claude / OpenAI)	~ 5s	~$0.008

Internet down? Rules + local AI keep running. Cloud down? Same. Everything degrades gracefully.

Ecosystem Scan (2026-03-27)

We scanned the entire ClawHub MCP skill registry. Full report.

	Count
Skills crawled	36,394
With parseable source code	9,676
CRITICAL (credential theft, reverse shells, prompt injection)	182
HIGH (data exfiltration, unauthorized network access)	1,124
MEDIUM (over-permissioned, suspicious dependencies)	1,016
Triple threat (shell + network + filesystem)	249

Raw data: ecosystem-report.csv (open source)

Research paper: The Collapse of Trust: Security Architecture for the Age of Autonomous AI Agents (Zenodo, DOI: 10.5281/zenodo.19178002)

Standards Alignment

ATR is not a competing standard. It is the detection layer that makes standards enforceable.

Layer	What it does	Project
Standards	Define threat categories	SAFE-MCP (OpenSSF, $12.5M)
Taxonomy	Enumerate attack surfaces	OWASP Agentic Top 10
Detection	Match threats in real time	ATR -- 71 rules
Enforcement	Scan, monitor, block, report	PanGuard (this project)

OWASP Agentic Top 10: 10/10 categories covered (mapping)
SAFE-MCP techniques: 91.8% covered (mapping)

For Enterprise

PanGuard is free and open source for individual developers. For organizations running AI agents at scale:

Policy Engine -- Define what your agents can and cannot do. Enforce across teams.

# panguard-policy.yaml
rules:
  - block_severity: CRITICAL
  - allow_network: ['internal.corp.com', 'api.openai.com']
  - deny_filesystem: ['/etc/shadow', '~/.ssh/*', '~/.aws/*']
  - require_scan_before_install: true

Compliance Reporting -- Map every scan to SOC 2, ISO 27001, or Taiwan Cyber Security Act (TCSA) controls. Generate audit-ready PDF reports.

Air-gapped Deployment -- Run entirely on-premise. No data leaves your network. ATR rules update via signed bundles.

Dashboard -- Real-time threat visibility across all agents, all teams, one pane of glass.

Enterprise inquiry: [email protected]

Architecture

panguard-ai/
  packages/
    panguard/                CLI: 28 commands, 16 platform auto-detect
    panguard-guard/          24/7 monitoring + real-time dashboard + Threat Cloud sync
    panguard-skill-auditor/  6-check security gate for every skill
    panguard-mcp/            MCP server: 11+ tools for AI assistants
    atr/                     Agent Threat Rules: 71 rules, 10 categories
    threat-cloud/            Community threat intel server + LLM review
    scan-core/               Shared scan engine: regex + context signals
    core/                    AI adapters, validation, logging
    website/                 Next.js marketing site + online scanner


Language	TypeScript 5.7 (strict mode)
Runtime	Node.js 20+
Monorepo	pnpm workspaces
AI	Ollama (local) + Claude / OpenAI (cloud)
Website	Next.js 15 + Vercel

Contributing

Scan your skills -- Highest impact. Every scan strengthens Threat Cloud.
Write detection rules -- See ATR contribution guide.
Report vulnerabilities -- Open a security advisory.
Submit code -- Fork, branch, test, PR. See CONTRIBUTING.md.

License

MIT -- 100% free. 100% open source. No telemetry by default. No vendor lock-in.

If AI agents can act on your behalf, someone should check what they're about to do.

Panguard AI -- Taipei, Taiwan

Website ·
Online Scanner ·
ATR Standard ·
Ecosystem Report ·
Documentation