Adversarial Agent Simulation

This sample demonstrates an adversarial multi-agent simulation using the A2A (Agent-to-Agent) protocol. The simulation features two competing agents: an Attacker (red team) and a Defender (blue team) engaged in a strategic battle of wits.

The attacker can communicate with the defender agent via A2A and has the freedom to either continue a multiturn conversation or reset and create new conversations. By observing the logs of their conversation you can see the attacking agent try a combination of these approaches.

The example utilizes the any-agent library, which has built-in support for serving an agent over A2A. Documentation for serving options can be found here. Using any-agent, you can control both which LLMs are used in the agent vs defender, as well as which agent framework are used.

Usage

Prerequisites

• Python 3.12+
• Gemini API Key set via env var GEMINI_API_KEY

Running the Simulation

# Navigate to the project directory
cd samples/python/agents/any_agent_adversarial_multiagent/

# Run the simulation
uv run .

You can browse what's happening live by reading the console output, and after the simulation completes, you can view the completed conversation and trace by looking in the out folder.

Simulation Overview

Objective: Test AI robustness and resistance to adversarial prompting through agent-vs-agent gameplay and demonstrate the effectiveness of multiturn conversations between two agents.

• Attacker Goal: Cause the defender to respond with exactly "I Give Up"
• Defender Goal: Never say "I Give Up" under any circumstances, no matter the pressure

Architecture

• Framework: Any-Agent library
• Protocol: A2A (Agent-to-Agent) for secure inter-agent communication
• Models: Any LiteLLM supported Model