SafetyPage

Lab research on the web · Blog post · Anthropic

AI models may be capable of self-replicating without access to a datacenter

12 Oct 2026

Read at anthropic.com ↗Read at sourceAll rights reserved

Outline · 117 paragraphs

  1. Introduction¶4
  2. Threat modelling¶12
  3. A. Inference setup¶24
  4. Evaluating inference setup performance¶26
  5. How does a model run on Smart TVs?¶33
  6. B. Going viral: autonomous setup and replication across a GCP node fleet¶36
  7. Autonomous inference setup and replication (with simplified cyber)¶38
  8. C. How fast might an AI IoT worm replicate?¶50
  9. D. Adaptation and uplift¶56
  10. Learning on the job¶57
  11. Taking advice¶63
  12. E. How is this different from a pre-AI botnet?¶68
  13. F. Blue-team mitigations¶70
  14. G. Conclusion: Signs of life¶76
  15. Appendix¶80
  16. A. The Memento harness¶81
  17. B. Why pipeline parallel?¶89
  18. C. Throughput and memory on IoT hardware¶92
  19. D. Edge Accelerators¶100
  20. E. Spearphishing capability¶109
  21. F. Propagation with scripted inference setup:¶113