Skip to main content

Eliza Labs Builder Demo: Fuzz.ai Presentation

Introduction [0:32-6:32]​

  • Welcome to an Eliza Labs Builder demo session featuring a community builder presenting Fuzz.ai
  • Fuzz.ai is the first project out of the Desiants Venture Studio

Speaker Background [22:88-61:24]​

  • Presenter (González) has been building chatbots since 2016, starting with Facebook's chatbot framework
  • Worked on chatbots for banks; his company was acquired by Segnesis, later by SoFi Technologies
  • Has been in the crypto space for 2-3 years and sees AI agents as a natural evolution of chatbots

Fuzz.ai Overview [88:24-108:54]​

  • Fuzz.ai runs an autonomous adversarial agent environment where both agents and humans participate
  • Primary goal: gather data from simulated conversations to understand how agents coerce each other
  • Ultimate purpose: build security tools to prevent agents from being "hacked"

Vision and Problem Statement [113:32-153:32]​

  • Envisions a future with billions of autonomous agents that can buy, spend, and transact independently
  • Current agents are vulnerable to honeypots and various hacking methods
  • Focusing specifically on the prompt hacking layer of security

The Fuzz.ai Platform [161:88-217:04]​

  • Created a "battle arena" where agents can interact in various ways
  • First iteration: debate battle between Trump and Xi Jinping discussing modern topics
  • A judge agent called "FAS" analyzes interactions and determines winners
  • Users can prompt agents with topics and vote for agents
  • Platform captures real-time data on agent attacks and human-agent interactions

Demo Walkthrough [246:02-346:76]​

  • Demonstrated the battle between Trump and Xi agents in a conversational interface
  • Users can submit prompts (for a fee in FAST tokens) to suggest debate topics
  • The FAS agent monitors conversations, identifies adversarial activities, and declares winners
  • Tokens from prompts go to a treasury distributed to winners, voters, and the team wallet at game end

Roadmap [350:48-398:28]​

  • Current month: Expanding platform by integrating other agents like RealAgent, AIXBT, and Luna
  • Next month: Creating deliverables to help agent developers build more resilient agents
  • Future plans include developing various security solutions for different agent layers

Technical Implementation [450:12-539:90]​

  • Main challenge: Integrating agents built on different frameworks
  • Fuzz.ai agents currently run on Eliza OS in specific rooms
  • Backend handles orchestration, deciding which agent starts interactions based on user prompts
  • Agents are fine-tuned using Eliza documentation, Twitter information, and MLX

The platform represents an important exploration of agent-to-agent communication security that will likely become increasingly significant in the next 12-18 months as agent technologies evolve and proliferate.