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

Phase · November 2026

AI for pentest reporting: the best model, and how to use it

The best AI for pentest reporting is DeepSeek V4.1 Flash, because a report is long input and long output, at $0.24 in and $0.90 out per million tokens.

Use Kimi K3.1 Pro when one finding needs a long narrative, and a local model when the notes cannot leave the machine.

Updated November 18, 2026Authorized work only
01Method

A finding is an asset, an observation, a consequence and a fix

Reporting is the last PTES stage and post-execution in NIST SP 800-115. The reader needs the asset, what was observed, why it matters, and what to change. CVSS is the severity vocabulary. Ask for the categories the notes support, not a guessed score. Leave exploit steps out.

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

Draft from the notes, then a person signs it

One finding per prompt. DeepSeek V4.1 Flash drafts the set. Move one hard finding to Kimi K3.1 Pro, whose max output is 943,718 tokens. If the notes must stay on the machine, draft on a local model.

Prompt

Turn these notes from an authorized test into one finding. Include the asset, what was observed, why it matters, the CVSS categories the notes support, and a fix. Do not invent observations. Do not include exploit code or replayable steps.

Silk Compute does not store your prompts or model outputs. The API choice is on the API ranking. Other grades are on the home matrix.