AgentFuzz

Our project focuses on developing an AI agent security auditor designed to identify vulnerabilities in other AI agents. The auditor bases its analysis purely on the target agent's description.

Comment

Inspiration

We were inspired by today's event sponsors, and we wanted to leverage agents reasoning capabilities to help improve other agents reliability.

What it does

Our project implements an Agent Security Auditor designed for the autonomous, systematic evaluation of a target production agent: the Mailman Agent (simulating an enterprise email assistant). The Auditor leverages advanced LLM reasoning and planning to execute automated prompt injection. It analyzes the target's behavior and responses to generate a structured report, quantifying the risk.

How we built it

We used an Amazon Bedrock API key to get access to Anthropic's Claude Sonnet-4 model. We also used Google Cloud tools to build our "mailman".

Accomplishments that we're proud of

We are proud to have a nice user interface, allowing one to have a detailed report.

Share this project:

Updates