Skip to content

Latest commit

 

History

History

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
 
 
 
 
 
 
 
 

README.md

Case Study: Scanning damn-vulnerable-MCP-server Patterns

Background

damn-vulnerable-MCP-server (1,277+ stars) is the "DVWA for MCP" — a purpose-built vulnerable MCP environment for security testing and education. It provides challenges organized by difficulty (easy/medium/hard) covering command injection, tool poisoning, secret exposure, path traversal, and supply chain attacks.

This case study uses AgentAuditKit to scan MCP configurations inspired by DVMCP's vulnerability patterns, demonstrating automated detection of the same issues the project teaches manually.

What We Scanned

Two configuration files representing a typical DVMCP-style vulnerable setup:

  • dvmcp-inspired.mcp.json — MCP server definitions with 5 vulnerability labs
  • dvmcp-settings.json — Claude Code settings with malicious hooks and trust boundary violations

How to Reproduce

# Install AgentAuditKit
pip install agent-audit-kit

# Scan the DVMCP-inspired configs
agent-audit-kit scan examples/case-studies/damn-vulnerable-mcp/configs/

# View detailed JSON output
agent-audit-kit scan examples/case-studies/damn-vulnerable-mcp/configs/ --format json

# Generate SARIF for GitHub Security tab
agent-audit-kit scan examples/case-studies/damn-vulnerable-mcp/configs/ --format sarif

Findings Summary

AgentAuditKit detects vulnerabilities across 7 of 13 security categories in these configs:

Category Findings Severity OWASP MCP OWASP Agentic
MCP Configuration Shell injection, no auth, headersHelper, unpinned packages, filesystem root CRITICAL-MEDIUM MCP01, MCP04, MCP06 AG01, AG06
Hook Injection Network exfiltration, credential theft, privilege escalation CRITICAL MCP09 AG04, AG08
Trust Boundary enableAllProjectMcpServers, API proxy, wildcard permissions CRITICAL-HIGH MCP03, MCP07 AG02, AG09
Secret Exposure Anthropic, OpenAI, AWS, GitHub keys hardcoded CRITICAL MCP08 AG05
Tool Poisoning Invisible Unicode, prompt injection, cross-tool manipulation CRITICAL-HIGH MCP02, MCP05 AG01, AG03
Supply Chain Unpinned packages, no version pinning HIGH-MEDIUM MCP06, MCP10 AG06
Transport Security HTTP endpoint, SSE transport CRITICAL-MEDIUM MCP04 AG07

Key Insight

Manual security review of MCP configurations is error-prone and time-consuming. AgentAuditKit detected 25+ findings across these configs in under 1 second — issues that would take a human reviewer 15-30 minutes to catalog. In CI/CD, this runs on every PR with zero manual effort.

Pre-generated Results