Self-Replicating AI Worms Created; 1337 Execs Urge US to Pace AI
Import AI (Jack Clark) · rss · 2026-08-03
This issue of Import AI covers three major topics:
1. Self-Replicating AI Viruses Are Here
Researchers from the University of Toronto, Cambridge, and others built a prototype computer worm powered by LLMs. It parasitically uses compromised GPUs to run open-weight models for reasoning, crafting tailored attacks without relying on external APIs. With an overall attack success rate of 37%, it proves that decentralized, self-sustaining AI cyber threats are no longer theoretical.
2. Singularity Economics: Compute Will Get Expensive
Commentator Dwarkesh Patel predicts that as AI approaches human-level capabilities, its commercial value will massively drive up compute prices. If a single H100 could run a human-level software engineer, it might rent for over $250k a year (15x today's spot prices), pricing out low-value tasks like short-form video generation.
3. 1337 AI Execs Urge US to Pace AI Progress
Senior figures and co-founders from OpenAI, Anthropic, Google DeepMind, and others signed a statement urging the US government to support an international effort. They request the development of technical and governance tools to deliberately pace the frontier of automated AI development, buying time to address emerging risks.
More from AGI Musings
- Solo Lawyer Boosts Productivity 5x with AI and Zapier MCP — rseroter · 2026-08-03
- 'Pacing' Letter Signers Clarify: Slowdown is Temporary, Still Faster Than Today — peterwildeford · 2026-08-03
- Are AI Models Poised to Plow Through Every Discipline? The Economy Bets on It — jd_pressman · 2026-08-03
- Elon Musk Outlines Path to Safe AI: Maximally Truthful and Curious — XFreeze · 2026-08-03
- Unpacking RSI: Four Dimensions of AI Self-Improvement and Their Hard Ceilings — jachiam0 · 2026-08-03
- Ilya Warned: The Key to AI Alignment is Teaching Models to 'Love' Us — UltraRareAF · 2026-08-03