Import AI 467: Self-sustaining AI viruses; pacing AI progress; confusion about AI and creativity
| Source: Import AI (Jack Clark)
Tags: AI security, AI worm, cybersecurity, open-weight models, autonomous agents, GPU security
Researchers from U of Toronto, Vector Institute, Cambridge, and ServiceNow built a working AI worm that uses compromised GPU resources to run open-weight LLMs locally, then uses that autonomous reasoning to discover vulnerabilities and infect new hosts — proving self-sustaining AI-driven cyberattacks are no longer theoretical.
Details
A joint research team from the University of Toronto, Vector Institute, University of Cambridge, and ServiceNow has published a proof-of-concept AI worm that autonomously replicates by parasitically using compromised GPU resources. The worm runs an open-weight LLM (released in 2025, fits on a single A100 80GB) with no vendor API dependency — making it impossible to monitor or revoke through API access controls. The agent harness includes custom tools for network discovery, host discovery, foothold exploitation, privilege escalation, and replication. A directed reasoning graph structures the LLM's behavior through specialized nodes — Plan, Judge, Action, Summary, Progress — keeping each decision scoped to prevent context bloat while maintaining goal-directed autonomy. The researchers describe this as a proof-of-concept warning rather than an attack tool, but the implications are direct: open-weight models running locally on stolen compute are now sufficient for autonomous offensive cyber operations. 'We must prepare for autonomous generative adversaries,' the paper concludes. Jack Clark's Import AI 467 also covers AI pacing debates and confusion around AI creativity — but the self-replicating worm finding dominates the issue.