Researchers published “AI Agents Enable Adaptive Computer Worms,” demonstrating a prototype AI-driven worm powered by an open-weight LLM running locally. In a controlled isolated network the worm propagated across heterogeneous Linux, Windows, and IoT devices by using recursive reasoning to detect and exploit whatever vulnerabilities each target presented.
Unlike classic worms that rely on a single fixed exploit, this agent adapts in real time. Compromised machines become part of the worm’s own infrastructure, supplying compute (especially GPUs) for further reasoning. The design requires no commercial AI service, rendering centralized safety controls irrelevant.
The researchers highlight three impacts: a qualitative shift in threat capability, the collapse of the traditional economic barrier (marginal cost approaches zero), and the need for new evaluation methods across open- and closed-weight models. Steve notes that the experiment remains academic for now because few devices yet have the local inference capacity required, and criminal groups currently prefer targeted extortion over uncontrolled worms.