AI Agents Create Self-Replicating Computer Worms; OpenAI Solves Open Math Problems
A prototype AI-based computer virus has been developed by researchers from the University of Toronto, Vector Institute, University of Cambridge, and ServiceNow. This self-replicating worm can compromise systems with a success rate of approximately 37% for complete attacks. It functions independently of vendor APIs, utilizing a single local GPU and a directed graph for its reasoning process. The worm demonstrates success rates of around 80% in identifying vulnerabilities, 53% in exploitation, and 88% in self-replication, benefiting from its decentralized structure. In other news, Dwarkesh Patel forecasts a potential 15x increase in AI computing costs, while over 1,300 employees from AI labs are calling for U.S. government backing for regulated AI development. Research indicates that AI agents face challenges in conducting original research, whereas OpenAI's Astra model has successfully addressed ten major issues in mathematics and computer science.
Key facts
- Researchers from University of Toronto, Vector Institute, University of Cambridge, and ServiceNow built a self-replicating AI worm.
- The worm uses open-weight LLMs on compromised GPUs to reason and attack new hosts.
- Success rates: ~80% vulnerability detection, ~53% exploitation, 88% self-replication, ~37% overall.
- The worm operates without vendor APIs, using only a single local GPU.
- Dwarkesh Patel predicts compute prices could rise 15x due to AI's increased monetization.
- Over 1,300 employees from major AI labs signed a statement asking US government to support pacing AI development.
- Shadow evaluation research found AI agents fail to produce original research at top ML conference level.
- OpenAI's Astra model solved ten open problems in math and computer science.
- Henry Yuen, Columbia professor, praised the significance of OpenAI's proofs.
- The 'Tech Tales' story explores AI context windows and mortality.
Entities
Institutions
- University of Toronto
- Vector Institute
- University of Cambridge
- ServiceNow
- OpenAI
- Anthropic
- Google DeepMind
- Thinking Machines
- Meta
- Safe Superintelligence Inc
- Princeton University
- Cornflower Labs
- UK AI Security Institute
- UC Berkeley
- Georgetown University (CSET)
- Johns Hopkins University
- Golden Gate Institute for AI
- AI Digest
- Stanford University
- Columbia University
Locations
- United States
- United Kingdom
- Canada