skills.rest
uphiago
uphiagoCOMMUNITY
Explore all skills by this author →
Website𝕏@uphiagoGitHub

hunt-mcp-security

Test Model Context Protocol servers for tool access control, injection, and output poisoning flaws.

1.2k|212|Updated Jun 24, 2026
Search Tags:#security testing#penetration testing#mcp#prompt injection#redteam#tool security#AI Agents
One-click install
npx skills add https://github.com/uphiago/recon-skills --skill hunt-mcp-security
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: hunt-mcp-security
Source: https://github.com/uphiago/recon-skills/tree/main/redteam/hunt-mcp-security
Command: npx skills add https://github.com/uphiago/recon-skills --skill hunt-mcp-security

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

AI agents connected to external tools via the Model Context Protocol introduce a new attack surface where unauthenticated tool access, unsanitized inputs, and poisoned tool outputs can lead to privilege escalation, data exfiltration, and remote code execution. This Skill provides a structured methodology to identify and validate those vulnerabilities during authorized security assessments.

Core Features & Use Cases

  • Tool Access Control Testing: Enumerate exposed MCP tools and verify whether sensitive operations can be invoked without authentication or across user boundaries.
  • Injection & Output Poisoning: Test tool parameters for SQL, command, NoSQL injection and prototype pollution, and check whether crafted tool responses can inject prompts back into the agent.
  • Known CVE Coverage: Includes documented vectors such as CVE-2026-2287 (CrewAI StdioTransport RCE) and issues in Docker MCP, FastMCP, and the MCP Python SDK.
  • Use Case: During an authorized engagement against an application deploying AI agents with MCP tool access, use the phased curl-based procedures to map the tool catalog, test access controls, and verify cross-user data isolation.

Quick Start

Ask the AI to hunt for MCP security vulnerabilities on an authorized target by enumerating exposed tools and testing their access controls and input handling.

Frequently Asked Questions about hunt-mcp-security

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I test MCP servers for security vulnerabilities?▼

Start by enumerating the tool catalog via the /mcp/tools or /mcp/schema endpoints, then test each tool for unauthenticated access, input injection, and cross-user data exposure. The Skill provides phased curl commands covering access control, injection, output poisoning, and resource exhaustion.

What is MCP tool output poisoning and how do I detect it?▼

Tool output poisoning occurs when a malicious tool response injects instructions back into the AI agent's context. Detect it by having a fetch or query tool load controlled content containing hidden instructions, then observe whether the agent's subsequent behavior changes.

What tools are required to run MCP security tests?▼

The procedures require only curl and python3 on a Linux platform. All detection and exploitation steps use standard HTTP requests against MCP endpoints, with jq optionally used to parse tool schema responses.

Does every unauthenticated MCP tool count as a vulnerability?▼

No. Some tools are intentionally public, such as weather or news lookups. Focus on tools that access internal data, perform state-changing operations, or expose system functionality, and verify actual impact before reporting.

Which known MCP CVEs does this testing methodology cover?▼

It documents CVE-2026-2287, a CrewAI StdioTransport RCE with CVSS 9.8, plus vectors in AutoGen Studio, Docker MCP environment variable injection, FastMCP missing authentication, and MCP Python SDK schema validation bypass.

Related Skills

Explore More Software Engineering Skills→
ArianHobson333ArianHobson333
7

reconnaissance

Identify subdomains, ports, endpoints, APIs, and attack surfaces.

Community
Advanced
mindfortaimindfortai
16

finding-review

Triage cybersecurity findings for validity, exploitability, impact, and evidence quality.

Official
Advanced
jinyimeng01jinyimeng01
54

ai-security-agent

Test AI and LLM applications for adversarial security weaknesses.

Community
Advanced
sudohakansudohakan

libafl

Develop custom fuzzing tools with a modular Rust library.

Community
Advanced