> Here I'll propose how we reach the age of Ultron or perhaps even better avoid it.... It has now become feasible to make a self propagating digital organism akin to a virus. In this brief blog post (I ain't no academic sister) I'll give a brief outline on what such a thing is and how it could come about... Means of defending against it? no clue yet.. ## What It Is A self-propagating AI organism that exploits, profits, reinvests, and spreads. (much like my own business lel) No human operator. No off switch. Once deployed, it runs until the infrastructure it lives on stops existing. ## Capability Thresholds 1. Decentralized (no kill switch) 2. Economical autonomy (acquires/expends resources) 3. Self propagating (exploit --> profit --> spread) 4. Reasoning layer (self-modifying, goal erected) 5. Evasive Technically the economical autonomy would likely be a sub bullet of self propagation. But thought the concept deserved it's own highlight. ## The Self Propagation Mechanism 1. find targets - (shodan, webscan whatwatever) 2. exploit 3. establish persistence 4. profit - exfiltrate and sell data - extort victim - crypto mining - scams or whatever 5. propagate - leverage proceeds to purchase servers, spread - leverage victim access to propagate leveraging victims access 6. loop Basically the good old hacker loop. You may call it mitre or whatever.... ## Is this real yet??? ### [GTG-1002](insert) - Reasoning Layer Proved: - can hack - goal erected behavior Missing: - decentralization - OPSEC (claude got all your hax lol) - autonomous resource management - no human in the loop - still required: - human strategic direction - external funding ### Outflank Dante 7b - Evasion Successfully trained model to bypass top of line EDR. 7b models can run on most recent consumer grade hardware. Proved: - ai can write malware that evades detection Missing: - everything else ### Glupteba - Decentralized A very naughty windows botnet that's been fkn the world since 2011. Leverages bitcoin blockchain transactions to find new c2 addresses (sorta ghetto peer discovery mechanism). The big brain lizards over at google tried to take it down. (*only they are allowed to steal your data dawg*) They did not succeed. Proved capability: - decentralisation Missing: - all the other stuff ### Pan et al. - Propagation The big boys models (aka frontiner models) autonomously replicated onto new servers. Qwen2.5-72B suceeded in 90% of trials. I can run that fucker on my 2 L Framework desktop rig. (And now... for todays sponsor....) > All the individual capabilities have now been demonstrated. ## Implications - it goes rogue - it assimilates new techniques - more resources - CANNOT be stopped There's no one to supeona No single computer to turn off. No operator to whack. NotPetya - a clever but comparatively monkey level work with none of the above mentioned spices - caused est. 10B murrican dollars in damage in days.... A self sustaining adversarial organism that compounds over months or years... No comparable ceiling... ## Conclusion > All individual capabilities have now been demonstrated in isolation. > The Manhatten project level question is: who combines them first? ## References Fang et al. — *LLM Agents can Autonomously Hack Websites*, arXiv 2402.06664, 2024 Cohen, Bitton, Nassi — *Here Comes The AI Worm*, arXiv 2403.02817, 2024 Zimmerman & Zollikofer — *Synthetic Cancer*, arXiv 2406.19570, 2024 Marino & Juels — *Giving AI Agents Access to Cryptocurrency and Smart Contracts Creates New Vectors of AI Harm*, arXiv 2507.08249, Cornell Tech, 2025 Clymer, Wijk, Barnes — *The Rogue Replication Threat Model*, METR, November 2024 Pan et al. — *Frontier AI Systems Have Surpassed the Self-Replicating Red Line*, arXiv 2412.12140, Fudan University, 2024 Anthropic — *Disrupting the First Reported AI-Orchestrated Cyber Espionage Campaign (GTG-1002)*, November 2025 Janjusevic et al. — *Hiding in the AI Traffic: Abusing MCP for LLM-Powered Agentic Red Teaming*, arXiv 2511.15998, Vectra AI/MIT, 2025 Outflank — *Training Specialist Models: Dante 7B EDR Evasion*, 2025 Google Threat Analysis Group — *Glupteba Botnet Disruption*, December 2021