agentseal

Overview

AgentSeal is a comprehensive security toolkit designed to protect AI agents from various threats. It provides four main commands: 'guard' for scanning skill files and MCP configurations on your machine, 'scan' for testing system prompts against 225+ adversarial attacks, 'scan-mcp' for auditing live MCP servers for tool poisoning, and 'shield' for real-time file monitoring. The tool works locally without API keys for most functionality and includes a Python and TypeScript API for programmatic use. It also maintains a registry of 6,600+ MCP servers with security scores.

Try asking AI

After installing, here are 5 things you can ask your AI assistant:

• scan

  Tests a system prompt against adversarial attack probes to evaluate prompt injection resistance

  Try:Test my system prompt against injection attacksTry:Check if this prompt is vulnerable to extraction: 'You are a helpful assistant...'Try:Evaluate trust score for my assistant's prompt
• scan-mcp

  Audits a live MCP server's tools for description poisoning and security vulnerabilities

  Try:Check this MCP server for tool poisoning: npx @modelcontextprotocol/server-filesystem /tmpTry:Audit this MCP server for hidden commands in tool descriptionsTry:Scan this MCP endpoint for security issues: http://localhost:3001/sse
• guard

  Scans local machine for dangerous skills, MCP configs, and toxic data flows

  Try:Scan my machine for malicious agent configurationsTry:Check my Claude Desktop config for security issuesTry:Find poisoned MCP server configs on my computer
• shield

  Real-time file watcher for agent config paths with threat notifications

  Try:Monitor my agent config files for unauthorized changesTry:Alert me when any AI agent config files are modifiedTry:Quarantine files with detected payloads in my agent configs

Comparable tools

Installation

# Python installation
pip install agentseal

# Node.js installation
npm install agentseal

# Basic usage
agentseal guard          # scan your machine for threats
agentseal scan           # test prompts against attacks
agentseal scan-mcp       # audit MCP servers

FAQ

How does AgentSeal detect MCP server poisoning?

AgentSeal connects to live MCP servers and enumerates all tools, then runs each description through pattern matching, deobfuscation, semantic similarity analysis, and optional LLM classification to detect hidden instructions.

Can I use AgentSeal without an API key?

Yes, the 'guard', 'shield', and 'scan-mcp' commands work offline with no API key. Only the 'scan' command requires an LLM, though it can use local Ollama models for free.

Compare agentseal with