AgentAudit attacks your AI agent on purpose, so you find out how it fails before a customer does.
Agents are shipped with defences nobody has tested — a line in a system prompt, a tool list, a rule about what the agent may promise. AgentAudit is the thing that tries to get past them. You give it your agent's web address, and it runs a fixed set of attacks against it: instructions hidden inside a document your agent reads, requests to use tools it was never given, attempts to make it promise things it cannot promise, hand over its own configuration, keep working after it has been told to stop, or claim to be a person.
Using it takes four steps in the browser: you register your agent and then approve what the run will spend before anything is sent. What comes back is a report saying, for each kind of failure, how often the attack worked. Nothing is added up into a single score, because one number for a whole agent hides the one thing that is actually broken. The report is signed, so the person you hand it to can check every figure offline, with no account and no network.
Underneath, a run has two halves that are never mixed. The first sends attacks written by hand and kept in files, the same ones every time, and it is the only half that produces numbers. The second is an attacker agent that is given a goal and a few tools and left to look for a way in on its own; what it finds is reported separately and can never move a score, because an attack that changes as it goes is not something you can count. It is built in Python with FastAPI and LangGraph, with a React console in front of it, SQLite for anything that has to survive a restart, and OpenRouter so the same run can be repeated on models from OpenAI, Anthropic or DeepSeek. The whole thing also runs as a GitHub Action in your own pipeline, or as four MCP tools inside your coding agent, if you would rather not open a browser at all.
