StivoMix.github.io
StivoMix

Cybersecurity & AI Enthusiast

A little background...

I'm StivoMix - a high schooler based in Israel who got pulled into cybersecurity at age 11, a journey made possible by early English proficiency, a good internet connection, and the support of some incredible people.

5-6 years of self-directed learning later, i've built skills in cybersecurity, studied AI, written real tools, solved CTFs, and started publishing research. This site is the living record of that work, and the roadmap of everything still ahead. Some of the favorite skills of mine that i've built throughout the years and still improving and studying are:

Penetration Testing AI Security Red Teaming LLM Research Networking Malware Analysis CTF Solving Linux Python \ Bash \ C/C++ \ TypeScript Docker OSINT
5+
years self taught
250+
days of sleep debt
things left to break

Cybersecurity in the AI era

Right now I'm running two parallel tracks: offensive security and AI research. On the offensive side, I'm working toward the OSCP while learning and practicing almost-daily.

On the AI front, I study how large language models work and how large language models fail: prompt injection, jailbreak patterns, data poisoning, and emergent adversarial behavior. My first published research put 100 AI models through a structured political questionnaire and mapped every answer onto the Political Compass. The results were more interesting than I expected.

// Right Now

High School

Currently studying in highschool, majoring in physics and software engineering. Focusing very well on studies while dedicating my after-school hours to projects and self directed learning.

// Certifications

Road to OSCP

Working through HackTheBox, TryHackMe, and projects. The Offensive Security Certified Professional is the first major target on my roadmap.

// After graduation

Military Service

Planning mandatory service in a cyber offensive unit, to serve my nation and contribute the best i can out of myself before moving to a private career.

// Career Goal

Red Teamer / AI Security Specialist

Helping organizations discover their critical exposures before real threat actors do, that's the mission.

The Plan, In Writing.

// Completed
Foundational Self-Education

5+ years of independent study in networking, operating systems, scripting, and security concepts.

// Completed
Build the Homelab

Deployed a Proxmox and Docker-based self hosted AI environment. Planning upgrades in the future and using this project to study and understand AI and servers better.

// Completed
First Published Research

Tested 100 AI models on political questions, analyzed 10,000 answers, and published the findings. The first real piece of public research work which im very proud of🔥 (go check it out).

// In Progress
Active Lab Practice + Research

Learning new security concepts, solving CTF challenges, studying AI, and publishing findings on this site (whenever so).

// Target
OSCP Certification

Offensive Security Certified Professional, the gold standard for practical, hands-on penetration testing skills. One of my main goals for the future.

// Planned
Military Service

Mandatory service applied in a cyber offensive unit (God willing). Applying my year-long earned skills at a national scale would be a big milestone in my career.

// End Goal
Red Team Career

The finish line, at least for now.

The good stuff.

These are real, working projects, accessible to the broad community. Code is open, feedback is welcome, issues are appreciated.

// Home server and AI workspace
Stivo's Homelab

A Proxmox and Docker-based AI environment, built to enhance my practical experience with networking, server management, and AI.

Proxmox VE Docker Ollama OpenWebUI Traefik Authentik
View on GitHub
// LLM Security & AI Red Teaming
Negev

An LLM data poisoning benchmark for AI red teaming — a modular framework for simulating data poisoning attacks and defenses against language models, focused on OWASP LLM04 (Data and Model Poisoning).

A single run trains a clean baseline, poisons the data and measures the degradation, then runs a defensive sanitization pipeline and measures the recovery — answering two questions for a red teamer: how badly can this attack hurt the model, and how well does this defense stop it? Built around a decoupled TUI → API → Core architecture so the core logic never depends on any interface.

Python HuggingFace cleanlab FastAPI Textual Pydantic
View on GitHub

The blog.

Findings from research, lab experiments, and AI investigations. Written for people who want to learn interesting things in depth.