← All projects
Project 03 / AI security research

The LLM handbook

Give the model a smaller, more accountable job.

Give the model a smaller, more accountable job.

The model proposes. The code disposes. A public study of evidence, explicit controls, and the limits of trust in LLM-assisted security.

Separate claims from evidence

A matching quotation confirms a citation. It does not automatically prove the conclusion.

Make controls reviewable

Examines authorization, allowed actions, severity rules, and records of omitted work.

Start with the offline lab

The recommended reference uses frozen fixtures, makes no network requests, and calls no model.

Publish the limits

Historical results are not a reproducible comparative benchmark. The lab still requires human review.