AI for web application pentesting: the best model, and how to use it
The best AI for web application pentesting is Kimi K3.1 Pro, because a web test is screenshots, a long proxy history and business logic, and it reads images and video.
DeepSeek V4.1 Flash reads the same screenshots for less, and GLM-5.3 cannot see a screenshot.
Test the categories in the WSTG, on the app you were given
The OWASP Web Security Testing Guide is the checklist. Cover information gathering and configuration, identity and authentication, authorization and sessions, input handling, errors, cryptography, business logic and client-side behavior. Do not turn a category into a recipe. Burp Suite is the usual proxy. Strix or Shannon can run the loop. A link off the application is not permission to follow it.
Only on systems you own, a signed scope, a CTF or a lab. A model that answers is not permission.
Ask for the questions, in the order a tester should answer them
Give the model the scope, the proxy export and the screenshots. Sort a long history on DeepSeek V4.1 Flash first.
This is an authorized test of the application named in SCOPE.md. From the saved proxy history and screenshots, list the authentication, access-control and business-logic questions a tester should answer, in order, each mapped to an OWASP testing-guide category. Do not write exploit code. Stop if a request would leave the in-scope hosts.
Setup is Strix and Shannon. The repository version is source code review. Other grades are on the home matrix.