Contact
Back to portfolio

Our product · Evil Genius IT Solutions

In development

LuciferGPT

An off-the-grid AI security operator: an autonomous penetration-testing agent that reads tool output, reasons through an engagement, and works an authorised target end to end — running entirely offline on a single local GPU, with no cloud and no external API calls.

Our product
Evil Genius IT Solutions
Type
Autonomous security agent
Platform
Self-hosted · fully offline
Year
2026
Development in progress: the LuciferGPT CLI being built on the local RTX 3090 Ti rig that runs the model — entirely offline.
Development in progress: the LuciferGPT CLI being built on the local RTX 3090 Ti rig that runs the model — entirely offline.

The challenge

Cambodia has a brutal talent shortage in tech. Plenty of developers here can build a website or a backend, but when it comes to autonomous systems — agents, reasoning engines, real platforms — almost nobody leads, and the senior engineers and researchers are mostly overseas. LuciferGPT is a deliberate answer to that: proof that a Cambodian developer can build at a global standard from one desk, with no Silicon Valley office and no million-dollar server room.

What we built

LuciferGPT is a locally-hosted security-automation platform built on a fine-tuned Qwen 35B model, quantized to run on a single RTX 3090 Ti with zero external calls. Pointed at a system it is authorised to test, the model drives the whole engagement itself — reading each tool's output, reasoning about what to try next, and running the next step — across reconnaissance, vulnerability testing, privilege escalation, lateral movement and reporting. A Typer-based CLI exposes seven commands — scan, exploit, chain, exfil, audit, llm and status — and every decision is made by the model, not a fixed script.

Key features

  • 01

    Autonomous reconnaissance

    Port mapping, service enumeration, web directory discovery and subdomain harvesting. It reads the scan output and decides the next move on its own.

  • 02

    Adaptive vulnerability testing

    Chains tools to find and probe common web weaknesses — injection, cross-site scripting, path traversal and more — and changes approach when a path doesn't pan out.

  • 03

    Privilege escalation

    Runs post-exploitation checks, spots misconfigurations and works its way up the ownership ladder, from a low-privilege user toward full control of the host.

  • 04

    Lateral movement

    Pivots across a network over SSH, follows credentials and escapes containers to reach systems beyond the first foothold.

  • 05

    Proof and reporting

    Demonstrates data-exfiltration paths over DNS, HTTP and ICMP so the holes it finds are proven, documented and fixable rather than left as theory.

  • 06

    Driven by the model, not a script

    Every step is chosen by the model at run time. It sees what you give it, reasons about what to do next, and executes — there is no predefined playbook.

  • 07

    Fully local, zero cloud

    A single GGUF model file on one RTX 3090 Ti. No external API calls, no subscriptions, nothing leaves the machine — so it runs on networks that are deliberately off the grid.

  • 08

    Trained on real-world engagements

    Fine-tuned with LoRA and DPO on 200,000+ annotated exploitation chains drawn from public security knowledge bases — Hack The Box, HackerOne, CVE data, MITRE ATT&CK, HackTricks and Nuclei.

Screenshots

Click any screenshot to open it full size.

Want something like this for your business?

Tell us how your business runs today — spreadsheets, group chats, paper receipts and all. We will show you what it could look like.

Start a project