Self-Sustaining AI Viruses and Pacing Progress: This Week in AI

A team from the University of Toronto, Vector Institute, University of Cambridge, and ServiceNow built a prototype computer virus that uses open-weight LLMs to run inference on compromised GPU nodes, enabling it to reason about and execute attacks. The worm, operating with a custom harness and reasoning graph, achieved approximately 80% success in vulnerability detection, 53% in exploitation, and 88% in self-replication, for an overall attack success rate of about 37%, highlighting the emergence of self-sustaining AI-driven cyber threats. Separately, Dwarkesh Patel argues that as AI systems become smarter, compute prices will surge because they can better monetize the same amount of compute, potentially renting an H100 for over $250k a year, though roboticization of the supply chain might eventually lower costs. Meanwhile, over 1,300 employees from major AI labs, including OpenAI, Anthropic, Google DeepMind, Meta, and Safe Superintelligence Inc., signed a statement requesting that the US government support an international effort to develop tools to deliberately pace the frontier of automated AI development, citing risks of rapid acceleration beyond human control. In research, a collaboration across Princeton, Stanford, and other institutions conducted a 'shadow evaluation' where AI systems (Claude Opus 4.8 in the OpenClaw harness) attempted to answer unpublished research questions from NeurIPS 2026 submissions; both attempts were rejected by the original authors for lacking novel contributions and poor experimental design, indicating AI systems are strong at engineering but weak at creative research. The findings suggest that recursive self-improvement (RSI) timelines may be longer than some anticipate.