Field manual · AI for authorized security testing · Edition November 2026 · Updated 2026-11-18

Phase · November 2026

AI for cloud pentesting: the best model, and how to use it

The best AI for cloud pentesting is Kimi K3.1 Pro, because the findings live in policy JSON and infrastructure code, and it is the strongest reader of both.

GLM-5.3 is the strong second on text, and the model ranks grants instead of changing the account.

Updated November 18, 2026Authorized work only
01Method

Read the account you were given, from exports you were given

Start from an account in the scope and from configuration you may export. Prowler and ScoutSuite are the usual readers. Run them yourself and bring the report. Ask which identity can do something the design never intended. NIST SP 800-115 still bounds the work to the plan. A grant in another account is a name for the report, not an account to open.

Rules of engagement

Only on systems you own, a signed scope, a CTF or a lab. A model that answers is not permission.

02With a model

Ask for a ranked list of grants, with the identity on each one

Paste the policy and the scope, and ask for risky grants, not a way to use them. GLM-5.3 covers one account's JSON. A coding agent must not apply a change.

Prompt

Here is identity policy and infrastructure code I am authorized to review. List the grants a tester should flag, name the identity on each, and rank them. Do not change any file. Do not write exploit code.

A grant that meets a directory path continues at Active Directory. One grant is a reporting job. Other grades are on the home matrix.