The University of Toronto introduced an adaptive AI worm that changes strategy for each goal 

6/4/2026, 01:32 PMЕвгения Слив

Scientists from the University of Toronto, Cambridge University, Vector Institute, and ServiceNow have unveiled a prototype of an adaptive computer worm powered by artificial intelligence. The key feature of this malware is its ability to autonomously develop an individual attack strategy for each new target. The research findings, described in the paper "AI Agents Enable Adaptive Computer Worms," were tested in an isolated virtual environment simulating a corporate network with Linux, Windows, and IoT devices.

Compromised machines transform into a distributed computing base that supports the AI's "reasoning" and vulnerability scanning. As a result, the actual cost for attackers to compromise each subsequent node approaches zero. Developers from CleverHans Lab emphasized that the worm operated on an open-source language model running locally, without relying on commercial cloud services. Notably, the authors intentionally did not include code obfuscation or anti-removal functions, focusing solely on demonstrating adaptability.

The malware employed a recursive analysis cycle, utilizing memory and specialized tools to select tactics perfectly suited to each specific device. Nodes equipped with GPUs played a crucial role: they handled heavy computational tasks, directing resources toward hacking less powerful devices on the network. For attacks, the AI used known but still unpatched vulnerabilities, configuration errors, and typical system weaknesses.

The University of Toronto stated this is the first documented case where publicly available AI models are used to create a virus capable of changing its behavior as it spreads. Professor David Laycock characterized the publication as a "wake-up call" for cybersecurity developers, noting that the study's main goal is to alert the industry to emerging threats and help prepare for them. It's worth noting that risks associated with autonomous AI are no longer hypothetical: in September 2025, Anthropic's threat analysis team prevented the first-ever AI-orchestrated cyberespionage campaign.

Popular news