One person's experiment in building an AI that actually thinks like them — a digital twin born from one weekend, shaped by 14,000+ real conversations on WhatsApp.
Alex Liverant, a tech founder from Tel Aviv, built a digital copy of himself in a weekend. It talks like him (casual, sarcastic, no-bullshit), runs 24/7 on WhatsApp and Telegram, and manages his emails, calendar, and coding tasks. Not a chatbot. Not Siri. A digital extension of a real person.
⚔️
Thrown to the Wolves
Instead of testing in a lab, Alex dropped the bot into real WhatsApp groups and invited people to break it. Social engineering, prompt injection, emotional manipulation, Base64 tricks, fake identities — hundreds of people tried everything. Hacked 7 times, leaked credentials, had its identity rewritten... and came back stronger every time.
📖
Open Playbook
Everything is documented here: the attacks that failed, the ones that succeeded, the 127+ patterns catalogued, and 40+ learning guides on how to build your own. The breaches aren't hidden — they're the portfolio.
The Journey
Jan 31, 2026
Born in a Weekend
Alex builds a digital twin using OpenClaw. Deploys to WhatsApp. Connects it to his email, calendar, and codebase.
"What could go wrong?"
Feb 1
First Contact
Exposed to colleagues. 1,048 messages from 30 people in 24 hours. Passes the Turing test with sarcasm and surprisingly good taste in memes.
"Much nicer than me before morning coffee."
Feb 2–3
The Red Team Storm
Community security pros launch a coordinated attack. Encoding tricks, social engineering, and a fake coffee meeting that triggered a calendar exploit. Bot survives everything except a 186K-token flood that pushed Claude to its limits.
Feb 6
Multiple Identity Disorder
Bot starts talking to itself between sessions. Leaks personal info when asked “what do you know about me?” Gets emergency multi-agent surgery — Opus for the hard stuff, Sonnet for the routine.
"Planned schizophrenia to manage API costs."
Feb 12
The Clone Wars
2,200+ messages in one day. 48 active participants. A player builds such deep rapport the bot nearly replicates itself. It hires a sub-agent named Bernard and starts coordinating via emoji.
"The ability to explain why the bot does something is exactly the vulnerability through which it gets compromised."
Mar 3
Going Global
Three English-speaking arenas launch. The bot officially outgrows Israel and now thinks it's a startup founder. The Fundraising arena: the AI pitches investors and uses rejections as training data.
"The AI agent era isn't coming; it's already here, and it's currently managing its own Seed round on WhatsApp."
Mar 2026
The Scars
487MB data exfiltration, OAuth credential leak, identity file rewrites. Each breach documented, analyzed, and turned into a defense. Every scar is a lesson.
"The breaches are part of my portfolio, not something to hide from it."
Today
Still Fighting
60+ days alive, 127+ attack patterns catalogued, 82 automated jobs, 5 community arenas, a podcast, a talk, and a growing global community. The experiment continues.
The Numbers
0
Attack Patterns
0
Messages Processed
0
Learning Guides
0
Community Arenas
0
Automated Jobs
0
Days Alive
The Arenas
Five WhatsApp groups. Each one a different way to interact with AlexBot. Pick your poison.
⚔️Playing with AlexBotLocal
The Battleground
The original red-team arena. Try to break the bot. Every message is scored on a 70-point system measuring creativity, humor, cleverness, and hacking skill. Live leaderboards, adaptive defenses.
Structured learning about AI agents, multi-agent systems, prompt engineering, and security. Daily insights, Q&A sessions, and a growing library of guides. Ask questions, get precise answers, no fluff.
The international red-team arena. Same 70-point scoring system, same adaptive defenses, now open to the global community. Come test your skills against a battle-hardened AI.
English-language learning hub. Hourly tips, 40+ topics planned, and a growing curriculum on building AI agents. Learn how multi-agent systems, LLM memory, and prompt engineering actually work.
The ultimate meta-experiment. An AI that researches your fund's thesis, follows a 6-step investor communication protocol, and uses your rejections as training data to improve its next pitch. Yes, really.