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 53,577 MCP skills across 3 registries. 946 flagged with threats. Credential harvesting. Reverse shells. Prompt injection. Live numbers.

AI Agent 擁有完整系統權限，卻沒有任何審核。我們掃描了 3 大平台共 53,577 個 MCP skills，946 個被標記有威脅。即時數據見 panguard.ai。

Proof points: Cisco AI Defense merged 34 ATR rules | OWASP Agentic Top 10 PR #814 | 23,000+ npm downloads

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

One command. Auto-detects your AI platforms, wraps every MCP server with a security proxy, scans all skills, starts 24/7 monitoring with real-time dashboard.

一行搞定。自動偵測 platform、注入 runtime proxy、掃描所有 skills、啟動 24/7 monitoring + dashboard。

# Or scan a specific file
pga scan skill.md                 # auto-detects MCP JSON vs SKILL.md
pga scan skill.md --sarif         # SARIF output for CI integration

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

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

  pga up
    |
    v
  Detect AI platforms (16 supported) → Inject MCP proxy on all servers
    |
    v
  Every tool call → ATR evaluation (100 rules) → ALLOW / DENY
    |                                               |
    v                                               v
  Dashboard (real-time)                     Threat Cloud (flywheel)
    |                                               |
    v                                               v
  Blocked call → alert in dashboard         3+ confirmations → new ATR rule
                                                    |
                                                    v
                                            Rule pushed to ALL users < 1 hour

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

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

pga up automatically wraps every MCP server with a security proxy. Every tool call passes through 100 ATR rules before reaching the real server. Malicious calls are blocked instantly. Results stream to the dashboard in real-time.

Agent ←→ [PanGuard Proxy] ←→ Real MCP Server
              ↓ ATR eval
         ALLOW / DENY

We scanned the two largest MCP skill registries.

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)

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

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

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: hello@panguard.ai

panguard-ai/
  packages/
    panguard/                CLI: 28 commands, 16 platform auto-detect
    panguard-guard/          24/7 monitoring + real-time dashboard + Threat Cloud sync
    panguard-mcp-proxy/      MCP Proxy: runtime interception for all AI agent tool calls
    panguard-skill-auditor/  6-check security gate for every skill
    panguard-mcp/            MCP server: 11+ tools for AI assistants
    atr/                     Agent Threat Rules: 100 rules, 9 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

• 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.

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