urd

redasgard/urd

★ 0 stars Python AI/LLM Updated today

Compositional trust analysis for Model Context Protocol (MCP) deployments. Detects authority paths that exist in practice but not in any individual server's manifest — where low-privilege tool output influences high-privilege tool execution without crossing a reviewed boundary.

View on GitHub → 🔍 Audit Wallet Slippage →

Quick Install

Copy the config for your editor. Some servers may need additional setup — check the README.

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "urd": {
      "command": "uvx",
      "args": [
        "urd"
      ]
    }
  }
}

Or install with pip: pip install urd

README Excerpt

**Breaking MCP Trust Boundaries: Cross-Server Authority Injection in Agent Toolchains** Urd is a research harness for compositional trust analysis in MCP deployments. This workshop repo demonstrates one concrete failure class: low-trust MCP output crossing through host planning context, selecting a protected record, and reappearing as a high-trust destructive tool parameter.

Topics

agent-securityai-securitydefcondefcon34llm-securitymcpmode-conpythonred-teamruntime-tracingsecurity-researchstatic-analysistrust-analysis

urd

Quick Install

README Excerpt

Topics

Related AI/LLM Servers