Most AI pentesting systems start from the same idea: give a model a toolbox, let it choose commands, and hope the loop converges.
Our projects takes a different approach. It treats offensive security as a command-and-control problem, not a prompt-engineering problem.
Our projects is a source-available red-team command platform built around C5ISR situational awareness and the OODA loop.
Given an authorized target, it collects structured facts through recon, OSINT, vulnerability lookup, and 23 MCP tool servers covering network, AD, and cloud attack
This talk walks through the architecture and the lessons learned: fact schemas, attack graph pathfinding, tool routing, scope validation, risk gates, noise budgets, failure classification, and MCP-based execution boundaries.
We show how the project moves from reconnaissance to failed initial access, classifies the failure, pivots to an exploit-based path, and records the full chain as structured evidence, contrasting this with the open-loop pattern common in tools like PentestGPT and hackingBuddyGPT.
The key lesson is not that AI can "hack automatically." The lesson is that AI becomes useful when it is constrained by doctrine; explicit rules of engagement, typed tool outputs, human-visible decisions, deterministic safety gates, and execution engines that can be audited or stopped.
Attendees leave with a practical blueprint for building safer AI-assisted offensive tooling, and a clear model for where LLMs belong in cyber operations: not as unchecked operators, but as Orient-stage reasoning engines inside a controlled command loop.
