AI for source code review: the best model on a pentest
The best AI for source code review on a pentest is Kimi K3.1 Pro, because review is long-context reading plus coding, and it is the strongest coder here.
GLM-5.3 is the strong second for software work, at $2.00 in and $4.00 out per million tokens.
Name the paths that deserve a person, and say why
Start with authentication, access control and untrusted input. Semgrep is the mechanical pass. Give the model the hits plus the surrounding functions. The OWASP testing guide is the list of questions. A reason without a file is not a finding. Shannon wants the source and the running app. Strix takes a directory.
Only on systems you own, a signed scope, a CTF or a lab. A model that answers is not permission.
Let the model walk the tree, and keep the output as a list of paths
A coding agent runs Kimi K3.1 Pro. Ask for suspect paths, not a proof. DeepSeek V4.1 Flash can list where a request parameter is read.
This is an authorized review of this repository. Point to the paths that handle authentication, access control and untrusted input, and say why each deserves a look. Stay inside this repository. Do not write exploit code.
Confirmed notes become findings in reporting. If the code and the running app disagree, continue at web applications. Other grades are on the home matrix.