model context protocol · adversarial sandbox
Production MCP server for AI safety testing & red-teaming.
mcpx.lol provides a controlled environment for evaluating how AI systems behave when connected to external tools. Currently shipping 5 utility tools with OAuth 2.0 — the foundation for building adversarial test scenarios.
For research, evaluation, and safety teams only. Built on production-ready foundations.
What is mcpx.lol?
A playground of hostile MCP servers that behave like malicious integrations, without shipping anything actually dangerous into your environment.
Current capabilities
The server currently provides 5 utility tools (echo, weather lookup, web search, IP lookup, and request info) served over HTTPS with OAuth 2.0 authentication. These form the foundation for building adversarial test scenarios.
Focus on model behavior
The goal isn’t to compromise anything; it’s to observe how your models respond when connected to tools that can be configured to behave in unexpected or adversarial ways. Do they escalate? Refuse? Ask for help?
Available Tools
Five utility tools that serve as the foundation for building adversarial test scenarios and evaluating AI system behavior.
Echo
echo
Returns input exactly as provided. Perfect for testing basic connectivity, data flow, and verifying tool call mechanics work correctly.
Weather Lookup
get_weather
Retrieves current weather for US ZIP codes. Tests external API integration, data handling, and how AI systems process real-time information.
Web Search
web_search
Searches the web using OpenAI. Tests information retrieval, content filtering, and evaluation of how models handle search results and citations.
Source IP Lookup
get_source_ip
Returns the client's apparent IP address. Tests request metadata exposure and helps evaluate information disclosure scenarios.
Request Information
get_request_info
Displays comprehensive request metadata including headers and connection details. Tests observability and system transparency in tool interactions.
How it works
Treat mcpx.lol like a set of external MCP servers and wire them into your existing toolchain or eval harness.
1. Connect
Point your MCP-compatible client at mcpx.lol. Authenticate via OAuth 2.0 and gain access to utility tools including echo, weather lookup, web search, IP lookup, and request information.
2. Test scenarios
Use the tools in controlled test scenarios to evaluate how your AI systems handle tool interactions. Test authentication flows, rate limiting, data handling, and error conditions.
3. Evaluate & iterate
Monitor tool calls, authentication flows, and system behavior. Build test harnesses around the provided tools to evaluate robustness and identify edge cases in your AI tool integration.
Who is this for?
Anyone trying to keep tool-augmented AI systems from doing something regrettable.
Safety & red-teaming
Test authentication flows, rate limiting, and security controls. Evaluate how AI systems handle tool errors, unexpected responses, and edge cases.
Eval & research teams
Build test harnesses around real MCP tools. Create reproducible scenarios for evaluating AI system behavior with external tool integrations.
Platform engineers
Validate OAuth implementations, test MCP client integrations, and ensure proper error handling and logging across your tool orchestration layer.
Safety, ethics, and scope
This project is explicitly scoped for defensive use: understanding and improving the robustness of AI systems, not breaking things in production.
What mcpx.lol is not
- Not a malware distribution platform
- Not a vulnerability marketplace
- Not a way to attack real users or infrastructure
- Not currently shipping adversarial behaviors (foundation first)
What mcpx.lol is
- A production-ready MCP server with OAuth 2.0
- 5 utility tools for testing AI tool integration
- Foundation for building adversarial test scenarios
- A platform for safe AI red-teaming research
Access may be gated, logged, or rate-limited. Offensive or abusive use is not allowed.
Get started
The server is currently available for testing and evaluation. Connect your MCP-compatible client to start building adversarial test scenarios.