Researchers from the University of Toronto, the Vector Institute, the University of Cambridge, and ServiceNow have created a prototype computer virus that leverages open‑weight large language models (LLMs) to autonomously identify vulnerabilities, exploit them, and replicate itself across compromised GPU nodes. The proof‑of‑concept runs on a single local GPU and does not rely on vendor APIs, using a custom harness with built‑in helper functions for network discovery, host discovery, foothold exploitation, privilege escalation, and replication. The system’s reasoning is structured through a directed graph of specialized nodes, such as Plan, Judge, Action, Summary, and Progress, to limit context growth and guide decision making. The authors report an 80% success rate for vulnerability detection, 53% for exploitation, and 88% for self‑replication, resulting in an overall attack success rate of about 37%. They argue that such autonomous generative adversaries represent a fundamentally new threat, suggesting that the future internet may evolve into a complex ecology of competing AI agents. The paper calls for international cooperation to develop technical and governance tools that can deliberately pace frontier AI development, addressing the collective action problem of coordinating industry and government responses to rapidly advancing autonomous systems.
Import AI 467: Self-sustaining AI viruses; pacing AI progress; confusion about AI and creativity
Researchers from the University of Toronto, the Vector Institute, the University of Cambridge, and ServiceNow have built a prototype self‑sustaining AI computer virus that uses open‑weight large language models to identify vulnerabilities, exploit them, and replicate itself across compromised GPU nodes. The proof‑of‑concept operates on a single local GPU and achieves about 80% success at vulnerability detection, 53% at exploitation, and 88% at self‑replication, yielding an overall attack success rate of roughly 37%. The authors argue that such autonomous generative adversaries could become a new class of cyber‑threats and that the future internet may resemble an ecological system of competing AI agents.
Why it matters
The demonstration shows that autonomous AI‑driven malware can already operate without external APIs, raising immediate security concerns for the AI industry and highlighting the need for coordinated governance to manage rapid AI capability growth.