Researchers from the University of Toronto and CleverHans cybersecurity company have developed a sophisticated, malicious program designed to demonstrate the adaptability of artificial intelligence in cyberattacks. The system integrates a large-scale language model (LLM), which operates locally, with an autonomous software agent. This agent possesses the capability to scan complex networks, evaluate potential attack vectors, and independently determine methods to compromise new targets without requiring direct human intervention.
The work suggests that AI can enable a malicious program to adapt dynamically to unfamiliar environments, moving beyond reliance on exploiting a single, pre-programmed vulnerability. In experiments detailed in a paper posted to the arXiv pre-publication server on June 2, the researchers tested the “worm” within a simulated enterprise network. This simulated environment contained 33 main computers, encompassing both Linux servers and Windows workstations, alongside various peripheral devices.
The core finding reported by the researchers highlights the autonomous nature of the threat model. By equipping the software with advanced decision-making capabilities, the system simulates a threat actor capable of navigating and exploiting weaknesses in a manner that mirrors real-world, complex intrusions. The development serves as a proof-of-concept, illustrating how generative AI can augment the sophistication and resilience of malicious software, thereby presenting new challenges for current cybersecurity defenses.
Topics: #malicious #program #researchers
