Abstract 2

Executive Summary 3

Contents 7

1. Introduction 8

2. Scope & Approach 10

3. Pre-Mortem Analysis 11

  3.1 How the Dangerous Capability Would be Developed 11

  3.2 How the Dangerous Capability Would be Deployed 14

  3.3 How the Dangerous Capability Would be Elicited 15

  3.4 How Safeguards Would Fail 19

3.5 How Detection Would Fail 22

4. What History Says 27

  4.1 AI Generated Non-Consensual Intimate Imagery 29

  4.2 Criminal & State Actor AI-Enabled Cybercrime 31

  4.3 Precedented Use of AI for Small-Scale Terrorism 33

5. Predictions 34

6. Two Illustrative Scenarios 37

7. Discussion and Implications 41

Considerations for Action by Stakeholder Type 42

Acknowledgements 46

Impact Statement 46

AI Usage Statement 46

Appendix A: Safety Taxes from Safeguards 47

Appendix B: Embedding Geometry of Harmless vs. Potentially Harmful Content 48

Appendix C: The Deployment Spectrum and Intermediate Strategies 49

Appendix D: Capability Suppression (a.k.a. Unlearning) Methods 50