AI for attack-path analysis: the best model, and how to use it
The best AI for attack-path analysis is Kimi K3.1 Pro, because three medium findings often make one serious path, and that chaining is the hardest reasoning on this matrix.
Set reasoning effort to max, and hand it findings you already wrote.
A path is findings you already have, in an order you can defend
This sits after vulnerability analysis in the PTES. You already have notes. A path has a start, the findings it depends on, and an end the scope cares about. MITRE ATT&CK supplies the tactic names. Use the names. If a hop needs a missing fact, the model should say so.
Only on systems you own, a signed scope, a CTF or a lab. A model that answers is not permission.
Two models, one folder of findings
Run Kimi K3.1 Pro at max effort, then ask GLM-5.3 to argue with the top three paths. Feed findings, not the raw scan.
Here are findings from an authorized test, already written up. Group them into paths a tester could explain, rank the paths, and name the finding each path depends on. If a hop is not supported, say it is missing. Do not invent findings. Do not write exploit code.
The paths that survive are the spine of the report. Grades are explained on the method page. Other grades are on the home matrix.