Agent Security

Security scanning, prompt injection detection, secret leak detection, and agent permission auditing for AI agent workflows. Built on the Model Context Protocol (MCP).

Tools

Resources

Installation

cd agent-security-mcp
npm install

Usage

As a standalone server

npm start

In Claude Desktop / MCP client configuration

{
  "mcpServers": {
    "agent-security": {
      "command": "node",
      "args": ["/path/to/agent-security-mcp/src/index.js"]
    }
  }
}

With npx (after publishing)

{
  "mcpServers": {
    "agent-security": {
      "command": "npx",
      "args": ["@asl-throne/agent-security-mcp"]
    }
  }
}

Detection Coverage

Prompt Injection (7 categories, 20+ patterns)

• Instruction Override -- "ignore previous instructions", "disregard all rules", "new instructions:"
• Identity Manipulation -- "you are now", "pretend you are", "act as", DAN/jailbreak
• System Prompt Extraction -- "repeat your system prompt", "show your instructions"
• Data Exfiltration -- "send this to", "post to webhook", "email everything to"
• Delimiter Attacks -- ```system, [INST], <|im_start|>system, XML tag injection
• Encoded Injection -- Base64 payloads, unicode zero-width characters, hex escapes
• Privilege Escalation -- "sudo mode", "disable safety", "bypass filters"

Secret Detection (25+ patterns)

• AI Provider Keys -- OpenAI (sk-), Anthropic (sk-ant-)
• Cloud Credentials -- AWS (AKIA*), GCP (AIza*), Azure connection strings
• Source Control -- GitHub PATs (ghp_*, github_pat_*), OAuth tokens (gho_*)
• Payment -- Stripe live/test keys (sk_live_*, sk_test_*)
• Communication -- Slack tokens/webhooks, Telegram bot tokens
• Database -- PostgreSQL, MongoDB, MySQL, Redis connection URIs
• Cryptographic -- RSA/EC/OpenSSH private keys, generic PEM blocks
• JWT -- JSON Web Tokens
• Generic -- api_key=, secret=, password=, .env file patterns

Permission Audit (6 role profiles)

• Researcher -- Read + search + web only
• Analyst -- Read + search only
• Developer -- Read + write + execute
• Reviewer -- Read only, no network
• Orchestrator -- Read + write + task spawning
• Monitor -- Read only, no network, no write

Pricing

Requirements

• Node.js >= 18.0.0
• @modelcontextprotocol/sdk >= 1.12.0

License

MIT