About This Report  iii

Summary  v

Figures and Tables  x

CHAPTER 1

Introduction  1

CHAPTER 2

Research Approach  4

Methodology  4

Limitations  6

CHAPTER 3

Attacks Against AI Systems  8

Evasion  8

Extraction  8

Poisoning  9

Misalignment  9

Efficiency  9

CHAPTER 4

Enumerating AI Vulnerabilities  11

Schematic of the AI System  11

AI Vulnerabilities  12

CHAPTER 5

Characterizing the Risk of AI Vulnerabilities  15

Threat  16

Impact  18

Risk  20

CHAPTER 6

Mitigations for AI Vulnerabilities  25

Training and Fine-Tuning Data Mitigations  25

Tokenizer and Embedding Mitigations  26

Transformer Mitigations  26

Reward Function Mitigations  27

Context Window and RAG Mitigations  27

CHAPTER 7

Recommendations  29

Prioritize Security Controls at the Data Layers  29

Treat Certain AI Vulnerabilities as Structural Risks Rather Than as Patchable Bugs  29

Incorporate Component-Level Risk Assessments into AI System Design  29

Develop Advanced Exploitation Detection and Monitoring Capabilities for AI Systems  30

Develop Evaluation Methods That Reflect Stochastic Exploitation  30

Establish Criteria for Cataloging AI Vulnerabilities  30

CHAPTER 8

Conclusion  32

APPENDIX A

List of AI Vulnerabilities  33

Standard Listing of Vulnerabilities  33

List of Vulnerabilities in JSON Format  42

APPENDIX B

Explanation of Threat and Impact Scores  55

Abbreviations  60

References  61

About the Authors  68