A proof of concept for autonomous large language model viruses

A collaboration between the University of Toronto, the Vector Institute, the University of Cambridge, and ServiceNow has produced a functioning computer virus powered by an open-weight large language model. The agent, described as a self-sustaining cyber-threat, does not merely execute static code; it uses local GPU resources to run inference, allowing it to adapt its exploitation strategy to the specific defenses it encounters on a target machine.

The prototype operates through a custom harness equipped with a reasoning graph. This structure decomposes the agent’s logic into specialized nodes—including planning, action selection, and progress evaluation—to prevent context window saturation and maintain focus on high-level goals. According to the research findings, the model achieved an 80 percent success rate in vulnerability detection and a 53 percent rate in exploitation. Once a foothold is established, the agent attempts to replicate itself onto new hosts with an 88 percent success rate.

Because the system relies on an open-weight model published in 2025 and fits within the 80GB VRAM of a single NVIDIA A100, it operates independently of vendor APIs. This autonomy prevents centralized providers from revoking access or monitoring the attack's progression. The researchers suggest that the resilience of such agents comes from their ability to form a decentralized swarm, where multiple replicas explore diverse exploitation paths concurrently. This architecture implies a shift in network security toward an environment where defense may eventually require autonomous agents capable of identifying and neutralizing adversarial models at machine speed.