University of Toronto researchers demonstrated a new cybersecurity threat: an AI-powered worm that exploits computer vulnerabilities. Unlike traditional worms requiring skilled programmers, this prototype uses open-weight AI models to spread through networks without human intervention. The worm represents a significant advancement in automated cyber attacks.

The worm tailors attacks to different platforms, including Linux, Windows and IoT devices. It gathers data as it spreads, siphoning passwords and uncovering new vulnerabilities. Most concerning, it uses processing power from infected machines to fuel its reasoning and strategy. If a flaw is patched, the worm simply exploits other weaknesses to re-infect the same machine.

This threat became more real with Anthropic's Mythos model, which has uncovered over 10,000 flaws. While the prototype only exploits known vulnerabilities, researchers fear bad actors could adapt it to find new ones. "In an interconnected world, no system is immune to this threat," said lead author Nicolas Papernot.