AI Worms Exploit Vulnerabilities in Devices via Custom Attack Playbooks

• •AI worms can autonomously exploit vulnerabilities without an attacker server.
• •The worm compromised 62% of a 33-host network in just seven days.
• •A zero-click attack on WeChat allows account hijacking before calls are answered.

Researchers at the University of Toronto demonstrated a self-replicating AI worm that compromised 62% of a simulated 33-host network within seven days. This worm, powered by a free large language model, autonomously identifies vulnerabilities in devices such as laptops, phones, and IoT systems. It utilizes a zero-click attack method exploiting WeChat's VoIP stack, allowing it to hijack accounts before calls are answered. The worm can generate tailored exploits in real-time, including those for newly disclosed vulnerabilities like the CopyFail Linux kernel bug. The study highlights the adaptability of this malware, which does not rely on fixed scripts and can operate without an attacker-controlled server. The findings raise significant concerns about the security of connected devices and the potential for widespread exploitation. Current defenses against such threats remain inadequate.