Project 03 / AI security research
The LLM handbook
Build an agent whose decisions you can inspect.
Build an agent whose decisions you can inspect.
A public, step-by-step field manual for building bounded autonomous security agents, with runnable lessons, an offline lab, and an honest evidence register.
Learn the control boundary
The model proposes; the application owns authorization, action admission, evidence, and acceptance.
Build it step by step
Runnable lessons progress from a fixture run to a bounded model connection and an assembled application.
Start without a target
The default lab is offline, uses synthetic proof rules, and makes no network or model calls.
Read the limits
Historical case studies and current controls are kept distinct; findings still require human review.