What is mulot?

mulot is an agentic AI web pentesting tool that controls a real web browser to test websites for security weaknesses, the way a human penetration tester would. Instead of relying on the biggest, most expensive AI models, it is designed to work with open weights models like GLM-5.2, Gemma, or Qwen, running a real headless Chromium browser rather than just sending raw HTTP requests. The idea behind the project is that the tools an agent has access to matter as much as the model itself. Security professionals commonly use tools like Burp Suite, ZAP, or Fiddler for intercepting and replaying web traffic, and language models have seen a lot of writing about these tools during training. So mulot gives the model that same kind of toolkit rather than something unfamiliar. The project is split into two halves. The proxy half keeps a full history of HTTP traffic in a SQLite database that can be queried and replayed, lets the agent rebuild and reissue requests, run automated fuzzing against a marked point in a request, and run passive or active scans over captured traffic. Because it captures traffic through the browser itself, it can read HTTPS content without needing a separate interception certificate. The thinking half runs JavaScript directly inside the target web page, letting the agent write its own scripts for things like padding oracle attacks or time based SQL injection, plus a library of pre-written playbooks and wordlists built into the program that get loaded based on what technology stack the target uses. To use it, you build the Go binary that acts as an MCP server, then run a small Python script called agent.py that connects the tool to any OpenAI-compatible model provider, with presets for OpenRouter, local llama.cpp, and Zai. A handful of environment variables control things like the browser's user agent, whether it runs headless, and any upstream proxy to route traffic through. The author reports that GLM-5.2, running only through this toolkit, solved 87 percent of the OverTheWire Natas web security challenges and 73 percent of Root-Me's Web-Server challenges. The README includes a detailed results table listing which specific challenges were solved and which technique each one required, along with a short note on the ones that failed.