MCPSec
Formal verification and security analysis of the Model Context Protocol
The first ProVerif-based formal security model of the Model Context Protocol (MCP), along with practical attack implementations that validate the formal findings on real MCP environments.
Research Overview
- Formal Modeling - ProVerif models of MCP protocol flows (initialization, tool discovery, sampling, OAuth 2.1, multi-server)
- Threat Verification - Formal verification of four threat models:
- Prompt Injection
- Tool Poisoning
- Data Leakage
- Privilege Escalation
- Practical Attacks - 8 attack scenarios implemented with the MCP Python SDK
- Mitigations - Formal re-verification with proposed protocol-level defenses
Project Structure
formal/ ProVerif formal models (.pv/.pvl files)
attacks/ Python attack implementations and MCP servers
paper/ LaTeX paper and artifacts
results/ Experimental results (generated)
docs/ Literature survey and documentation
scripts/ Top-level automation
Significance
This project bridges my expertise in formal verification and protocol security with the emerging field of LLM security, demonstrating how traditional security analysis techniques can be applied to AI systems.