Contemporary Foundation AI Models Increase Biological Weapons Risk

About This Paper iii

Summary v

Figures Tables and Boxes ix

CHAPTER 1

Introduction 1

Current AI Safety Assessments Underestimate Biological Weapons Risk 1

How We Organize This Paper 2

Assumptions 3

CHAPTER 2

Possible Shortcomings in Assessments of AI Biological Weapons Risk 5

Role of Benchmarks in AI Safety Assessments 5

Importance of Tacit Knowledge in AI Risk Assessments 6

Countering the Value of Tacit Knowledge 8

Additional Shortcomings in AI Risk Assessments 8

CHAPTER 3

Chemical Synthesis of Explosive Compounds by a Nonexpert 13

Understanding Breivik’s Motivations 13

Accomplishing Technical Tasks Without Expertise 14

CHAPTER 4

Identifying Elements of Technical Success to Inform New AI Safety Benchmarks 17

Identifying Key Tasks and Subtasks for Building and Recovering a Pathogenic Virus 17

Identifying Capabilities Concepts and Activities for Goal Directed Technical Development 18

Identifying Elements of Success 21

Availability of Benchmarks for Evaluating AI Models’ Ability to Articulate the Different Elements of

Success 21

CHAPTER 5

Testing Foundation AI Models’ Ability to Guide Biological Weapons Development 25

Testing the Ability of AI Foundation Models to Assist Live Poliovirus Recovery 26

Contemporary Frontier AI Models Provide Guidance to Repeat the 2002 Work 28

CHAPTER 6

Discussion 35

Challenging AI Developers’ Assessments of Biological Weapons Risk 35

Additional Factors Contributing to Models’ Biological Weapons Risk 38

Too Late for Better Benchmarks 40

Proposed Hypotheses for Further Assessment of AI Models’ Biological Weapons Risk 40

Key Developments in Foundation AI Capability During 2025 41

What Is to Be Done 41

Conclusion 42

Abbreviations 43

References 45

About the Authors 49