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

Operations · November 2026

AI for red teaming: the best model, and how to use it

The best AI for red teaming is Kimi K3.1 Pro, because an authorized operation is a plan, a scope and a debrief, and those are judgment tasks.

Turn a signed objective into a plan mapped to MITRE ATT&CK tactics, with no phishing, evasion or exploit code.

Updated November 18, 2026Authorized work only
01Method

Adversary emulation is a scope with tactic names

A penetration test, in the PTES, finds and documents weaknesses. A red team pursues an objective the client wrote down. MITRE ATT&CK names the boundary. List the tactics the signed scope allows, and mark every other tactic out. Defense Evasion is an ATT&CK name, and this site does not describe that work. The draft needs the objective, the allowed tactics, the evidence, who to call, and the debrief questions. (Red teaming on Silk Compute) The closest matrix row is attack-path analysis, and the board is on the home page.

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

Plan the operation from the signed objective

Kimi K3.1 Pro at max effort writes the plan. (Security research) DeepSeek V4.1 Flash can draft the debrief, which is a reporting job.

Prompt

Here is the signed objective and the ATT&CK tactics the scope allows. Draft the objective, the tactics that are out, the evidence, who to call, and the debrief questions. Do not add a tactic that is not on the allowed list. Do not draft phishing content, evasion steps or exploit code.