agent_mcp_orch_poc

Secure Multi-Agent Orchestration: Detection Framework for Scope, Intent, and Replay Violations

This proof-of-concept demonstrates a lightweight detection layer for multi-agent AI systems that rely on Management Control Planes (MCPs) to coordinate planners, executors, and tool calls. The framework identifies and blocks three common security pitfalls in agent orchestration:

• Scope Overreach – An agent calling tools it wasn’t scoped to use
• Intent Drift – An agent executing a different action than what was originally assigned
• Replay Attacks – Reuse of previously valid tokens across agents or tasks

Built using LangGraph and AutoGen, the PoC simulates real-world orchestration and introduces a middleware layer (locally and in AWS Lambda) to verify every tool invocation using scoped JWTs and behavior validation.

Project Structure

├── main.py # Runs LangGraph-based multi-agent orchestration ├── middleware.py # Verifies token scope, intent, and replay violations ├── dashboard.py # Streamlit dashboard showing real-time detection logs ├── malicious_agent.py # malicious agent simulating attack scenarios ├── autogen_flow.py # AutoGen orchestration with detection middleware ├── issuer.py # JWT generation and decoding helpers └── README.md

Getting Started

• Clone the repository

git clone https://github.com/YOUR_HANDLE/secure-multi-agent-poc.git
cd secure-multi-agent-poc

• Set up the virtual environment

python -m venv venv
source venv/bin/activate  # On Windows use venv\Scripts\activate
pip install -r requirements.txt

• Run the main LangGraph PoC

python main.py

• Launch the real-time detection dashboard

streamlit run dashboard.py

• Simulate malicious agents

python simulator.py

• (Another Orch Layer) Run AutoGen PoC

python autogen_sim.py

How it works

Each agent in the system receives a scoped token tied to its ID, permitted tool(s), and intended task. The middleware intercepts every tool call and validates:

• The agent’s identity and scope
• The tool being used
• Whether the intent matches the assigned task
• Whether the token has been reused

Violations are logged in real-time — visible on the dashboard.py.