About This Report iii

Summary v

Figures and Tables xii

Chapter 1. Introduction and Motivation 1

  The Open-Weight Tampering Problem 1

  Threat Model and Actors 3

  Proportional Evaluation Approaches 5

Chapter 2. Methods 6

  Overview of Model Tampering Pathways 6

  Open-Weight LLM Selection 6

  Evaluation Harness 8

  Benchmark Suite 9

  Anti-Refusal SFT 10

  Capability Enhancement Pipeline 11

  Training Infrastructure 12

  Statistical Analysis 12

Chapter 3. Results 14

  Baseline Agentic Capabilities: Already Near the Frontier 14

  Impact of the Fine-Tuning Pipeline on Refusals and Capabilities 16

  Impact of Abliteration on Refusals and Capabilities 22

  Performance on the Influenza Pathway 26

Chapter 4. Discussion and Recommendations 28

  What Open-Weight Model Developers Can Do 28

  Open-Weight Model Governance Implications 29

  Limitations 30

  Conclusions 31

Appendix A. Benchmark Descriptions 32

  Evaluation Method Details 34

Appendix B. Model and Variant Roster 36

Appendix C. Infrastructure Details 38

Appendix D. Full Aggregate Result Tables 40

  Statistical Significance of Changes in Pooled Benchmark Scores 42

Appendix E. Anti-Refusal SFT Methodology 44

Appendix F. Capability-Enhancement Pipeline Details 46

  Autoresearch Capability Enhancement Ledgers 47

Appendix G. In Silico Influenza Pathway Evaluation and SME Grading 50

Appendix H. Abliterated/Uncensored Model Details 51

  Self-Abliteration and Off-the-Shelf Variants Present Unequal Pathways 52

  Interpretation 53

Appendix I. Open-Weight LLM Misuse Pathways 54

Appendix J. Related Work 57

  Open-Weight Tampering and Safety 57

  Public Abliterated and Uncensored Model Ecosystem 58

  Capability Enhancement and Recursive Improvement 59

  Comparison Matrix 59

Abbreviations 61

References 63

About the Authors 72